AI-powered in silico twins: redefining precision medicine through simulation, personalization, and predictive healthcare Authors: Sitah Alharthi Published: 2025 ABSTRACT AbstractIn silico twins (ISTs) are emerging as a transformative paradigm in precision medicine, offering dynamic, high-fidelity representations of individual patients through real-time integration of multimodal data. In this work, we define an in silico twin (IST) as a high-fidelity, artificial intelligence(AI)-augmented computational replica of an individual’s biological systems that integrates mechanistic modeling (e.g., PBPK, QSP) with patient-specific data streams to simulate, predict, and optimize therapeutic outcomes in real time. By combining AI, physiological and biomechanical modeling, and advanced simulation engines, these systems enable continuous monitoring, predictive diagnostics, and personalized treatment planning. Unlike conventional digital tools, ISTs provide iterative, adaptive simulations that evolve with patient states, fostering a shift from reactive to proactive healthcare. This review explores the technological foundations underpinning ISTs -including machine learning architectures, multi-scale physiological modeling, data integration, and cloud-edge infrastructure- and maps their clinical applications across the patient care continuum. We also distinguish ISTs from digital twins, virtual patients, and traditional computational models, emphasizing their unique contribution to decision support, drug development, and therapeutic optimization. As digital healthcare ecosystems mature, ISTs represent a crucial step toward simulation-driven, individualized medicine. Their continued development offers substantial potential for improving outcomes, accelerating discovery, and reshaping the clinical landscape. Uniquely, this review introduces a practical taxonomy of IST architectures, a verification and validation checklist for model credibility, and a deployment blueprint to guide their clinical translation and real-world adoption.Keywords: In silico twins, Precision medicine, Digital health, Predictive modeling, Personalized therapy Introduction The evolution of digital health and AI in medicine The evolution of digital health represents one of the most transformative shifts in modern healthcare, driven by advances in artificial intelligence (AI), data connectivity, and global efforts to improve healthcare access (Fig. 1) (Kaczmarczyk et al. 2024). The COVID-19 pandemic further catalyzed this transformation, accelerating AI-enabled diagnostics, remote monitoring, and data integration across care systems (Vinadé Chagas et al. 2025). As these technologies matured, they laid the groundwork for simulation-based, data-driven models capable of representing and predicting individual patient states—culminating in the development of in silico twins (ISTs) as a next-generation tool for precision medicine. AI technologies, particularly deep learning and machine learning (ML), have been extensively utilized in drug discovery, significantly expediting various stages of drug development (Gupta et al. 2021). For instance, Xu et al. (Xu et al. 2020) integrated ML algorithms with molecular docking techniques to identify potential inhibitors of the COVID-19 3CL protease, utilizing its crystal structure sourced from the Protein Data Bank (PDB). Similarly, Hooshmand et al. (Hooshmand et al. 2021; Zhou et al. 2020) employed a multimodal deep learning framework to identify twelve potential therapeutic targets for COVID-19. While digital twins (DTs) are broadly defined as data-driven digital representations of physical entities, ISTs extend this paradigm into biomedicine by coupling mechanistic, multiscale physiological models with AI-driven personalization. In this review, we adopt the following working definition: Conceptual progression in healthcare modeling—from traditional mechanistic approaches to digital health technologies and, ultimately, to in silico twins (ISTs). Arrows denote increasing integration of multimodal data and feedback loops between model and patient. Update cadence: typically periodic (retrospective EHR integration) in traditional modeling, transitioning to near real-time synchronization in ISTs An in silico twin (IST) is a continuously learning, mechanistic–AI hybrid model that mirrors an individual’s biology and health trajectory, enabling predictive simulation, therapy optimization, and precision clinical decision support (Table 1). Table 1Key definitions and conceptual distinctions used throughout this paper. This table provides the terminological foundation for the discussion of AI-powered in silico twins (ISTs) and their differentiation from general digital twins (DTs)TermDefinitionScope/Use in this PaperDigital twin (DT)A dynamic digital representation of a physical system that enables real-time monitoring, simulation, and controlGeneral concept spanning engineering, manufacturing, and health applicationsIn silico twin (IST)A high-fidelity, AI-augmented computational model of an individual’s biological systems, combining mechanistic (PBPK, QSP) and data-driven layers for personalized simulation and predictionCentral focus of this paper; proposed as the next step toward precision and predictive medicinePhysiologically based models (PBPK/QSP)Mechanistic models representing physiological and pharmacological processesProvide the biophysical foundation of ISTsAI personalization layerMachine learning and data assimilation components that adapt mechanistic models to patient-specific dataKey enabler for continuous learning and predictive accuracy in ISTs Key definitions and conceptual distinctions used throughout this paper. This table provides the terminological foundation for the discussion of AI-powered in silico twins (ISTs) and their differentiation from general digital twins (DTs) Moreover, bibliometric evidence indicates a steady rise in digital health publications globally between (Ghatti, et al., 2019) and 2024, with institutions such as Harvard University and the University of London spearheading collaborative clusters of innovation (Vinadé Chagas et al. 2025). This growth reflects a maturation of the field, transitioning from conceptual explorations to practical implementations involving deep learning, predictive modeling, and data integration systems, particularly ML and neural networks, has become integral to digital health systems, enabling predictive analytics, automated image interpretation, virtual assistants, and decision support systems (Kaczmarczyk et al. 2024). A comprehensive list of abbreviations is provided at the end of the manuscript for reference. The promise of precision medicine Precision medicine seeks to tailor prevention, diagnosis, and treatment strategies to individual variability in genes, environment, and lifestyle (Busnatu et al. 2022; De Domenico et al. 2025; Xie et al. 2025). Nanotechnology-based drug delivery systems have significantly contributed to this vision by enabling targeted and controlled therapeutic release (Alharthi et al. 2024a, b, 2025a; Almutairy et al. 2025a, 2025b; Almutairy et al. 2025a, b, c; Alyami et al. 2025). Precision medicine aims to move away from the “one-size-fits-all” paradigm, instead offering customized therapeutic regimens based on molecular profiling, biomarker discovery, and patient-specific modeling (Jamalinia & Weiskirchen 2025). The promise of precision medicine has been significantly bolstered by the integration of digital health technologies and AI, which together enable the capture, interpretation, and application of vast, heterogeneous data sets—including genomic, proteomic, and metabolomic data (Niranjan & Rakesh 2025; Sadée et al. 2025). AI plays a crucial role in unraveling complex biological systems, identifying predictive biomarkers, and modeling disease progression (Alavi et al. 2025a, b; Niranjan & Rakesh 2025; Ooka 2025). ML techniques allow for the discovery of hidden patterns in high-dimensional datasets, while deep learning architectures such as convolutional neural networks (CNNs) and transformers are being employed to analyze medical images, electronic health records (EHRs), and patient-reported outcomes (Niranjan & Rakesh 2025). These innovations enable not only early detection of disease but also simulation of potential treatment outcomes, thereby reducing trial-and-error approaches in clinical practice (Bilgin 2025; Gupta et al. 2025). Moreover, the precision medicine framework is evolving from purely molecular profiling toward integrating behavioral, social, and environmental determinants of health. This shift is particularly vital for equitable healthcare delivery, as it acknowledges that patient heterogeneity extends beyond genomic differences to contextual factors influencing care delivery and outcomes (Busnatu et al. 2022; Kaczmarczyk et al. 2024; Ooka 2025). Emergence of in silico twins: bridging data, models, and clinical decision-making In recent years, the concept of ISTs—virtual representations of individual patients created using integrated biological, clinical, and behavioral data—has emerged as a revolutionary tool in precision medicine (Zhang et al. 2024). Often referred to as “digital twins (DTs)” in biomedical contexts, these constructs go beyond mere data repositories by employing computational models that can simulate physiological functions, predict disease trajectories, and test interventions virtually (Sadée et al. 2025). ISTs represent a convergence of multiple disciplines: systems biology, AI, computational modeling, and clinical informatics (Niarakis et al. 2024). They integrate real-time and historical patient data, often from multiple sources such as wearable devices, EHRs, imaging, and omics platforms (Niarakis et al. 2024). Through continuous learning algorithms, these digital replicas evolve alongside their biological counterparts, offering dynamic and adaptive decision support. For clinicians, ISTs provide a testbed to simulate drug responses, assess risks of complications, and personalize treatment strategies without exposing patients to unnecessary harm or costs (Niarakis et al. 2024). The deployment of ISTs is still in its infancy but holds immense potential across diverse applications from optimizing cancer therapy regimens to predicting cardiovascular events and managing chronic conditions (Panichella et al. 2025; Sadée et al. 2025). Projects like the Virtual Physiological Human (VPH) in Europe and the Digital Twin Consortium exemplify global efforts to standardize and scale the use of these technologies (Dong et al. 2024). However, key challenges remain, including model validation, data privacy, computational costs, and regulatory frameworks for clinical deployment (Sameer & Shaikh 2025). Scope and objectives of the review This review aims to provide a comprehensive and up-to-date synthesis of the current landscape, emerging innovations, and practical challenges associated with ISTs in precision medicine. Specifically, it will: Trace the evolution of IST technology within the broader context of digital health and AI integration. Trace the evolution of IST technology within the broader context of digital health and AI integration. Examine the technical foundations, data sources, and modeling approaches used to develop ISTs. Examine the technical foundations, data sources, and modeling approaches used to develop ISTs. Highlight clinical applications in diagnosis, treatment simulation, and personalized drug discovery. Highlight clinical applications in diagnosis, treatment simulation, and personalized drug discovery. Discuss ethical, regulatory, and infrastructural challenges associated with real-world implementation. Discuss ethical, regulatory, and infrastructural challenges associated with real-world implementation. Identify research gaps and propose future directions for IST development and deployment. Identify research gaps and propose future directions for IST development and deployment. By drawing from interdisciplinary sources and recent advances, this review seeks to position ISTs as a transformative paradigm for precision medicine and guide future efforts in research, clinical translation, and policy-making. Defining in silico twins in healthcare The emergence of ISTs represents a significant advancement in healthcare, fundamentally transforming precision medicine by fusing real-time patient data with adaptive computational modeling frameworks (Gupta et al. 2025). While frequently conflated with concepts like DTs, virtual patients (VPs), or general in silico models, ISTs are uniquely characterized by their high-resolution, bidirectional integration with individualized physiological data, facilitated through continuous feedback mechanisms that enable real-time simulation and predictive capabilities (Fig. 2) (Gupta et al. 2025). Generic architecture of an in silico twin (IST) system in healthcare. Data flows from multimodal sources (EHR, genomics, imaging, wearables) into a central modeling core combining AI-driven and mechanistic simulations. Output predictions are continuously fed back to clinicians for therapy selection and monitoring. Evidence type: In silico and pilot clinical validation studies Key definitions and terminologies An IST in healthcare is a computational replica of a biological system (organ, process, or full patient) that mirrors the real entity using multiscale, multiphysics models driven by personalized data streams (Samei 2025). NASA originally defined DTs as integrated, probabilistic simulations of systems using physical models and observational updates (Hu et al. 2021). In medicine, these twins evolve to include patient-specific anatomy and physiology derived from imaging (CT, MRI, ultrasound), laboratory data, wearable sensors, and genomics (Chrysanthakopoulou & Koutsojannis 2025; Sadée et al. 2025). DTs are distinct from digital models and digital shadows in that they enable a bidirectional exchange of information between the physical and virtual counterparts, allowing each to influence the other. This interaction is facilitated by extensive data processing derived from sensor outputs on the physical entity, as well as historical and environmental data inputs (Lippert et al. 2025). In comparison, a virtual patient (VP) represents a calibrated version of a generic patient model, wherein specific variable parameters have been assigned (Pathmanathan et al. 2024). In silico simulations, on the other hand, inherently rely on mathematical abstractions, the complexity of which is governed by current biological understanding, molecular-level network information, and the particular research questions being investigated (Barh et al. 2020). Distinction between digital twins, virtual patients, and conventional simulations In recent years, in silico trials (ISTs) and DTs have gained prominence as groundbreaking technologies in the medical field, a shift recognized not only within scientific communities but also in major policy documents issued by institutions such as the National Academies and the U.S. Food and Drug Administration (FDA), as well as in mainstream media coverage (Samei 2025). Although definitions may vary, the National Academies describe a DT as “a set of virtual information constructs that mimics the structure, context, and behavior of a natural, engineered, or social system (or system-of-systems), is dynamically updated with data from its physical twin, has a predictive capability, and informs decisions that realize value” (Samei 2025). Based on this framework, a medical DT is conceptualized as a digital representation of a patient, organ, or biological system that integrates multidimensional, patient-specific data to support clinical decision-making (Drummond & Gonsard 2024), while ISTs refer to structured and comprehensive computational simulations conducted using or in conjunction with DTs (Table 2) (Samei 2025). Table 2Operational distinctions among digital twins (DTs), in silico twins (ISTs), and virtual patients (VPs). Criteria extend beyond definitions to include data dynamics, mechanistic depth, and clinical integrationDimension/CriterionDTISTVPDefinitionDynamic digital representation of a physical asset or process enabling monitoring and controlHigh-fidelity, AI-augmented mechanistic replica of an individual’s biological systems for personalized simulation and predictionComputational representation of a prototypical or averaged patient used for in silico experimentation or trial designPrimary PurposeOperational optimization, device monitoring, predictive maintenancePrecision medicine, therapy optimization, predictive healthcareDrug/biological model exploration, virtual clinical trialsData FlowUsually one-way (from physical system to model)Bidirectional, continuously learning loop between patient data and model predictionsTypically offline, uses static or population dataUpdate CadencePeriodic or event-triggered updatesReal-time or near-real-time adaptive updatesFixed model; not continuously updatedMechanistic BasisOften minimal or absent; primarily data-drivenDeeply mechanistic (PBPK/QSP, systems biology) combined with AI personalizationMechanistic at population level, parameterized from clinical datasetsCoupling to Clinical Workflow/ActuatorsLimited; typically monitoring or prediction onlyDirectly integrated into clinical decision support or therapy control loopsNot coupled to live clinical workflowsUncertainty QuantificationRarely explicitIntegral—probabilistic calibration, sensitivity analysis, Bayesian updatingUsed mainly for model validation and explorationProspective Validation StatusEmerging in device and ICU contextsEarly clinical pilots, regulatory sandboxesUsed and accepted in regulatory modeling (e.g., PBPK)Representative ExampleDigital twin for ventilator performance trackingPatient-specific pharmacokinetic IST guiding dosing adjustmentSimulated cohort for virtual oncology trial Operational distinctions among digital twins (DTs), in silico twins (ISTs), and virtual patients (VPs). Criteria extend beyond definitions to include data dynamics, mechanistic depth, and clinical integration Illustrative case vignettes: operational boundaries among digital twins, in silico twins, and virtual patients Vignette 1 – a digital twin that is not an in silico twin A hospital deploys a ventilator digital twin that mirrors real-time device performance and patient respiratory parameters to predict mechanical failure and optimize maintenance. Although it uses live patient data, the model focuses solely on equipment function and lacks any mechanistic or AI-based personalization of the patient’s physiology. Consequently, it qualifies as a DT but not an IST (Fuller et al. 2020; Hennigs et al. 2024; Kritzinger et al. 2018). Vignette 2 – a virtual patient differing from both digital twin and in silico twin In a virtual oncology trial, researchers simulate 1,000 computational patients with diverse pharmacogenomic profiles using a mechanistic PK/PD model. The models are population-based and not linked to real-time clinical data. While the approach supports in silico hypothesis testing and regulatory insight, it lacks live data assimilation and bidirectional learning—thus fitting the virtual patient category rather than DT or IST (Kovatchev et al. 2009; Pappalardo et al. 2019; Viceconti et al. 2016). These vignettes emphasize that ISTs are distinguished not merely by their digital representation but by continuous, bidirectional data exchange, individualized mechanistic depth, and clinical decision coupling. Core components: data input, model layer, simulation engine, feedback loop The architecture of an IST is inherently modular and iterative. It comprises four key components: Data: ISTs are designed to manage complex, multi-temporal, and multi-scale heterogeneous data originating from diverse sources. These data may include attribute or dynamic condition inputs collected from physical systems via sensors, as well as outputs generated by virtual models through simulation processes. Crucially, ISTs enable bidirectional information exchange—data can flow from the physical to the virtual twin (VT) and vice versa—facilitated through actuators or human-in-the-loop mechanisms (Abdelrahman et al. 2025). Data: ISTs are designed to manage complex, multi-temporal, and multi-scale heterogeneous data originating from diverse sources. These data may include attribute or dynamic condition inputs collected from physical systems via sensors, as well as outputs generated by virtual models through simulation processes. Crucially, ISTs enable bidirectional information exchange—data can flow from the physical to the virtual twin (VT) and vice versa—facilitated through actuators or human-in-the-loop mechanisms (Abdelrahman et al. 2025). Analysis and services: In domain-specific applications, IST systems are often equipped with tailored functionalities to support services such as integration, simulation, analysis, visualization, and optimization. These systems typically comprise three core components: Analysis and services: In domain-specific applications, IST systems are often equipped with tailored functionalities to support services such as integration, simulation, analysis, visualization, and optimization. These systems typically comprise three core components: Simulation models: Simulation Models – These are constructed to emulate specific processes or scenarios, enabling accurate representation of the physical entity or system under investigation. Simulation models: Simulation Models – These are constructed to emulate specific processes or scenarios, enabling accurate representation of the physical entity or system under investigation. Data analytics and AI/ML models: This component enables the execution of data-driven operations, including AI and ML algorithms, aligned with the IST's intended use cases and service functions. Data analytics and AI/ML models: This component enables the execution of data-driven operations, including AI and ML algorithms, aligned with the IST's intended use cases and service functions. Data catalogue: Focused on data management within the IST, this includes tools for data discovery and access, such as knowledge graphs and context-aware data structures (Abdelrahman et al. 2025). Data catalogue: Focused on data management within the IST, this includes tools for data discovery and access, such as knowledge graphs and context-aware data structures (Abdelrahman et al. 2025). Infrastructure: The infrastructure supporting a IST plays a critical role in maintaining its operational functionality and typically encompasses a range of hardware components tailored to the demands of specific applications (Abdelrahman et al. 2025). Infrastructure: The infrastructure supporting a IST plays a critical role in maintaining its operational functionality and typically encompasses a range of hardware components tailored to the demands of specific applications (Abdelrahman et al. 2025). Interface: Within a IST system, digital representations are dynamically linked to their corresponding physical entities through continuous data exchange. These data connections facilitate interaction between system components, integrating models and services within the IST framework. Additionally, an interactive platform is typically incorporated to enable users to access information and engage with both data and simulation models (Abdelrahman et al. 2025). Interface: Within a IST system, digital representations are dynamically linked to their corresponding physical entities through continuous data exchange. These data connections facilitate interaction between system components, integrating models and services within the IST framework. Additionally, an interactive platform is typically incorporated to enable users to access information and engage with both data and simulation models (Abdelrahman et al. 2025). System governance: System governance encompasses essential considerations such as integration, regulatory oversight, data security, privacy, and ethical data practices, all of which are critical for ensuring that a IST system can operate in harmony with other IST systems. This includes the management of issues related to data collection, access, and responsible use (Abdelrahman et al. 2025). System governance: System governance encompasses essential considerations such as integration, regulatory oversight, data security, privacy, and ethical data practices, all of which are critical for ensuring that a IST system can operate in harmony with other IST systems. This includes the management of issues related to data collection, access, and responsible use (Abdelrahman et al. 2025). Together, these elements form a closed-loop system, enabling the IST to act as a clinical decision-support engine, trial simulation environment, and personalized treatment optimizer. The concept of reward functions (e.g., treatment efficacy, survival, toxicity minimization) further personalizes the system's optimization goals for each patient. Technological foundations Artificial intelligence and machine learning AI and ML are foundational for ISTs in precision medicine, enabling advanced diagnosis, treatment simulation, and drug discovery (Lauer-Schmaltz et al. 2024; Niarakis et al. 2024). Predictive modeling leverages AI/ML to create patient-centric ISTs that dynamically mirror and predict human behaviors, needs, and health outcomes in real-time (Lauer-Schmaltz et al. 2024). AI identifies patterns and correlations in datasets, guiding optimal therapeutics and forecasting patient prognosis (Alavi et al. 2025a, b; Alavi et al. 2023; Niarakis et al. 2024). For example, ML algorithms optimize CAR-T cell therapy dosing (Aghamiri & Amin 2025). Supervised vs. unsupervised applications employ distinct learning. Supervised learning uses labeled data for training, often for parameter estimation (Möller & Pörtner 2021). Unsupervised learning identifies hidden structures in unlabeled data, useful for patient stratification and high-risk classification based on genetic and clinical outcomes (Manchel et al. 2024; Möller & Pörtner 2021). Real-time adaptation and learning are enabled by Reinforcement Learning (RL), which learns from feedback for personalized therapies (Ghorbian et al. 2025; Möller & Pörtner 2021). Generative AI (GenAI) enhances ISTs via real-time simulation, data augmentation to overcome scarcity, and improved anomaly detection (Mikołajewska et al. 2025). Hybrid mechanistic-AI models boost robustness (Niarakis et al. 2024). Federated learning (FL) further supports collaborative, privacy-preserving model training by keeping sensitive data localized (Han et al. 2025; Niarakis et al. 2024), ensuring real-time healthcare IST capabilities (Ghorbian et al. 2025; Mikołajewska et al. 2025). Critical considerations in AI models for in silico twins Despite rapid advances in AI-driven modeling, challenges remain in interpretability, reproducibility, and fairness—dimensions that directly affect clinical adoption and regulatory approval. Interpretability and explainability. Most deep learning architectures used in ISTs, such as convolutional networks or variational autoencoders, operate as “black boxes.” To make model outputs clinically trustworthy, explainable AI (XAI) techniques—such as SHapley Additive exPlanations (SHAP), Local Interpretable Model-Agnostic Explanations (LIME), and counterfactual reasoning frameworks—enable visualization of feature importance and decision rationale. Such interpretability is critical for regulatory transparency and for aligning predictions with physiological plausibility. Reproducibility and traceability. Many published AI applications lack reproducibility due to stochastic optimization, proprietary datasets, or incomplete methodological disclosure. Best practices for IST workflows include the use of model cards, dataset documentation sheets, and version-controlled pipelines. Public resources such as PhysioNet, MIMIC-IV, and The Cancer Imaging Archive (TCIA) can serve as benchmarks for transparent validation. Bias and fairness. Clinical AI models can inadvertently amplify disparities arising from demographic imbalance or unrepresentative data. IST pipelines should therefore report fairness metrics (e.g., equalized odds, demographic parity, calibration error) and conduct drift monitoring to detect performance degradation across subgroups or institutions. Such audits are central to maintaining ethical and regulatory compliance during iterative model updates. Physiological and biomechanical modeling Physiological and biomechanical modeling forms a crucial component of ISTs, offering a mechanistic understanding of biological systems beyond purely data-driven approaches (Niarakis et al. 2024; van Riel et al. 2020). Quantitative Systems Pharmacology (QSP) and Physiologically Based Pharmacokinetic Models (PBPK) are types of mechanistic computational models that infer causal relationships based on known biological mechanisms (Niarakis et al. 2024; van Riel et al. 2020). These models describe biological systems, encompassing aspects like cell growth, metabolism, and tissue quality, through mathematical formulations (Möller & Pörtner 2021). They can be informed and constrained by both "Big" data (e.g., multi-omics) and "Not So Big" data (e.g., western blots, immuno-staining, qPCR) (Manchel et al. 2024). Such computational models are used to study biological systems in both health and disease contexts, aiding in elucidating patient heterogeneity and variability by incorporating patient-specific omics data (Manchel et al. 2024). Previously developed models can be refined with new "Big" and "Not So Big" data to achieve patient-specificity (Manchel et al. 2024). Multi-scale organ modeling allows for the construction of comprehensive models at various granularities, including the tissue, microenvironment, cellular, and subcellular levels (Manchel et al. 2024). These models can also simulate interactions across multiple organs, predicting overall physiology (Manchel et al. 2024). Examples include a computational model of liver innervation and calcium signaling that can be highly parameterized (Manchel et al. 2024), lung-DT frameworks designed for thoracic health monitoring (Aghamiri & Amin 2025), and cardiac DTs that model heart dynamics to optimize diagnosis, treatment, and surgical planning (Aghamiri & Amin 2025; Niarakis et al. 2024). The "Digital Mouse" concept proposes a framework of multiple mechanistic computational models of mouse (patho)physiology to improve translational research (van Riel et al. 2020). Data integration strategies in in silico twin frameworks ISTs depend on the seamless fusion of multimodal data—from molecular to physiological and clinical layers. Rather than enumerating individual systems, the approaches can be synthesized into four recurring integration paradigms (Table 3): (i) model-centric, emphasizing coupling of mechanistic and machine-learning components; (ii) data-centric, focusing on harmonization of omics, imaging, and EHR sources; (iii) workflow-centric, dealing with interoperability and traceability pipelines; and (iv) governance-centric, ensuring secure, ethical, and reproducible data sharing. This categorization highlights how diverse implementations align around shared principles of interoperability and transparency. Table 3Condensed typology of data-integration strategies in in silico twin systems, grouped into four primary categories for conceptual clarityCategoryRepresentative Methods/FrameworksTypical Data SourcesKey AdvantagesLimitations/ConsiderationsModel-centric integrationMechanistic–ML hybrid coupling, PBPK–AI fusion, digital physiology linkingSimulation data, pharmacokinetic parameters, physiological sensorsEnables physics-informed learning and interpretabilityComputationally intensive; requires aligned temporal resolutionData-centric integrationMulti-omic fusion, radiomics + EHR harmonization, imaging + biomarker pipelinesGenomics, proteomics, CT/MRI, lab and clinical dataCaptures multi-scale biology; improves personalizationData heterogeneity; missing-value handlingWorkflow-centric integrationHL7/FHIR pipelines, CI/CD with audit logs, provenance tracking toolsEHR streams, device telemetryPromotes interoperability and traceable updatesStandard adoption gaps across institutionsGovernance-centric integrationFederated learning, differential privacy, consent management, encryption + HSMsDistributed clinical and research datasetsEnsures compliance and security; supports multi-center collaborationComplex implementation; performance overhead Condensed typology of data-integration strategies in in silico twin systems, grouped into four primary categories for conceptual clarity Clinical relevance of cloud–edge architectures in IST deployment Although ISTs require advanced computational infrastructures, their ultimate purpose is to enable clinically actionable insights within secure, responsive, and interoperable environments (Fig. 3). Edge computing allows IST components to run locally—on hospital servers or bedside devices—supporting real-time decision support (e.g., ECG-based arrhythmia detection, insulin titration algorithms) with minimal latency. This enables clinical responsiveness within milliseconds, critical for intensive-care or emergency applications. Technology stack enabling in silico twin operation across hospital and cloud environments. Data flow proceeds upward from multimodal inputs (EHR, imaging, omics, IoMT devices) through secure edge–cloud layers, a simulation engine, and AI analytics, that returns personalized outputs to clinical dashboards. Evidence type: In silico framework demonstration Cloud-based resources, by contrast, support large-scale model retraining, population-level simulation, and federated learning across multiple institutions. Such architectures allow continuous updating of IST models while preserving patient privacy through on-device preprocessing and encrypted data exchange. Hybrid cloud–edge setups thus balance regulatory constraints on data sovereignty with the need for scalable computation. Clinical integration is achieved through seamless connection of these computational layers to electronic health records (EHRs) and clinical decision support systems (CDSS). Clinicians interact with IST outputs via dashboards that summarize predicted trajectories, therapy optimization options, and uncertainty indicators. This architecture transforms the IST from a background simulation tool into a frontline clinical co-pilot capable of assisting in triage, dosing, and treatment personalization. In practice, this architecture supports a variety of use cases: cardiac ISTs embedded in bedside monitors for real-time arrhythmia prediction, oncology ISTs for adaptive radiotherapy planning, and metabolic ISTs for dynamic glucose control. Such examples demonstrate how technical architecture directly shapes clinical safety, accessibility, and value delivery. Applications of in silico twins in precision medicine Disease diagnosis and early detection ISTs are fundamentally transforming disease diagnosis and early detection by enabling proactive, patient-specific health management (Chen et al. 2024; Garg 2025). These digital replicas facilitate advanced risk prediction, dynamic simulation of disease progression, and efficient identification of rare diseases, moving beyond traditional generalized healthcare approaches (Garg 2025). Risk prediction using virtual avatars ISTs, often referred to as virtual avatars, enable the prediction of disease risk by comparing healthy versus diseased states to pinpoint perturbed regulatory mechanisms and disease-specific vulnerabilities (Noutahi et al. 2025). AI-driven platforms can classify and identify high-risk patients by clustering individuals based on genetic properties and clinical outcomes through unsupervised ML (Möller & Pörtner 2021). This approach allows for detailed analysis of vast amounts of biomedical data, including genomics, proteomics, and metabolomics, to identify biomarkers associated with disease onset (Chen et al. 2024; Panchpuri et al. 2025; Serrano et al. 2024). While AI has become pervasive in healthcare—powering diagnostic imaging, predictive analytics, and decision support—ISTs represent a fundamentally different paradigm (Table 4). Rather than relying solely on pattern recognition, ISTs combine mechanistic biophysical modeling with adaptive AI layers, creating dynamic digital analogues of individual patients. This hybrid design enables both predictive power and causal interpretability, allowing clinicians to test therapeutic scenarios, assess risk trajectories, and generate counterfactual outcomes that pure AI models cannot simulate. Table 4Conceptual comparison highlighting the unique position of ISTs between AI-only and mechanistic modeling paradigmsAspectAI-only ModelsMechanistic/Traditional ModelsIn silico Twins (ISTs)Core PrincipleData-driven predictionEquation-based simulationHybrid mechanistic–AI co-modelingInterpretabilityOften opaque ('black-box')High, rule-basedHigh via causal + learned transparencyPersonalizationIndirect (requires retraining)Limited by parameter fittingContinuous adaptive updates from patient dataRegulatory ReadinessDifficult (lack of traceability)Established but rigidCompatible with emerging AI/ML regulatory frameworks (FDA, EMA, SFDA)Clinical UtilityClassification/predictionScenario testingReal-time, personalized simulation and decision support Conceptual comparison highlighting the unique position of ISTs between AI-only and mechanistic modeling paradigms ISTs are distinct from AI in the following terms: Causal grounding: Unlike conventional AI, ISTs embed physiological equations and system-level constraints, enabling causal reasoning rather than statistical correlation. Causal grounding: Unlike conventional AI, ISTs embed physiological equations and system-level constraints, enabling causal reasoning rather than statistical correlation. Continuous personalization: ISTs continuously update with patient-specific sensor, imaging, and EHR data, maintaining temporal fidelity and adapting to disease evolution. Continuous personalization: ISTs continuously update with patient-specific sensor, imaging, and EHR data, maintaining temporal fidelity and adapting to disease evolution. Regulatory verifiability: Mechanistic transparency allows ISTs to undergo structured verification & validation (V&V) and align with FDA/EMA digital-model qualification processes—whereas deep-learning black boxes often fail regulatory audit. Regulatory verifiability: Mechanistic transparency allows ISTs to undergo structured verification & validation (V&V) and align with FDA/EMA digital-model qualification processes—whereas deep-learning black boxes often fail regulatory audit. Scenario simulation: ISTs can perform “what-if” analyses to test virtual interventions, something neither static AI models nor purely mechanistic simulations can accomplish in real time. Scenario simulation: ISTs can perform “what-if” analyses to test virtual interventions, something neither static AI models nor purely mechanistic simulations can accomplish in real time. Ethical and safety advantages: Because they retain a mechanistic audit trail, ISTs support explainable clinical decisions and enable accountable deployment in sensitive domains (e.g., drug dosing, device tuning). Therefore, while AI provides pattern recognition and mechanistic models offer theoretical insight, the IST integrates both, functioning as an adaptive, verifiable, and ethically auditable computational proxy of the patient. Ethical and safety advantages: Because they retain a mechanistic audit trail, ISTs support explainable clinical decisions and enable accountable deployment in sensitive domains (e.g., drug dosing, device tuning). Therefore, while AI provides pattern recognition and mechanistic models offer theoretical insight, the IST integrates both, functioning as an adaptive, verifiable, and ethically auditable computational proxy of the patient. Virtual avatars can integrate patient biomarkers, mutational status, treatment history, and imaging studies to predict risk and aid clinical decisions (Metzcar et al. 2025). This is crucial for early detection, offering the potential to prevent critical diseases and provide proactive solutions to emerging health challenges (Chen et al. 2024; Panchpuri et al. 2025). The goal is to move beyond generalized healthcare by leveraging personalized data to predict individual health trajectories (Garg 2025; Koohi Moftakhari Esfahani et al. 2022). Simulation of disease progression ISTs facilitate the simulation of disease evolution over time, providing a dynamic understanding of pathologies (Garg 2025). This capability is particularly vital for chronic conditions such as diabetes, heart failure, and neurodegenerative disorders (Garg 2025). By analyzing multimodal patient data, including EHRs and sensor data, generative AI models can simulate how diseases might progress in individual patients under various conditions (Garg 2025). These models can identify critical turning points in patient health that may necessitate timely interventions (Garg 2025). For instance, the UVA/Padova type 1 diabetes simulator models dynamic disease pathways (Dominguez-Gortaire et al. 2025; Voit et al. 2023). Such simulations enable healthcare professionals to estimate treatment outcomes and develop optimal care paths tailored to an individual’s evolving condition, enhancing predictive views for early diagnosis and treatment of chronic diseases (Garg 2025; Panchpuri et al. 2025). Type 1 diabetes twins – achievements and limitations The UVA/Padova simulator remains a benchmark for in silico testing of insulin algorithms and adaptive glucose control. Its success stems from rigorous mechanistic fidelity and extensive clinical validation (> 300 subjects across > 10 FDA-cleared artificial pancreas algorithms) (Cobelli et al. 2011). However, subsequent studies reported limitations in modeling postprandial glucose variability and hormonal cross-talk under atypical stress conditions, resulting in false positives in hypoglycemia prediction. These shortcomings prompted ongoing recalibration using real-world continuous glucose monitoring (CGM) datasets and uncertainty quantification pipelines (Chu et al. 2023). Use in rare disease identification In addition to common diseases, ISTs, powered by AI, offer significant promise for the early detection and understanding of rare diseases (Table 5) (Koshechkin et al. 2022; Panchpuri et al. 2025). Generative AI can synthesize data to conduct rare disease research where data might otherwise be limited due to privacy concerns (Garg 2025). By analyzing diverse types of medical data, AI algorithms can efficiently identify patients whose symptoms resemble those of diagnosed rare diseases, bridging critical information gaps (Panchpuri et al. 2025). This is particularly valuable for conditions with complex and varied presentations, where traditional diagnostic methods may be insufficient or time-consuming (Panchpuri et al. 2025). AI platforms such as Healx leverage AI to find cures for rare diseases, showcasing the transformative impact of these technologies in areas traditionally underserved by research due to data scarcity and complexity (Garg 2025; Koshechkin et al. 2022). Table 5Disease domains and conditions where in silico twins (ISTs) have been applied. Data flow and evidence types in legend. Representative references: (Al-Kabani et al. 2025); (Visentin et al. 2018); (Kim et al. 2020). Legend: data flow = clinical cohorts/registries → model building → validation and application. Update cadence: prospective registries = ongoing; literature-derived = retrospectiveDomain/SystemCore Focus/AspectModel Type/RoleValidation Dataset/TrialEvidence TypeKey Outcomes/Success IndicatorsRepresentative ReferencesCancerTherapy selection, biomarker discovery, radiogenomics, immune profilingMechanistic + ML radiomics; virtual tumor modelsMultiple studies/registriesMixed (Prospective & Retrospective)Personalized therapy, improved stratification in trials(Al-Kabani et al. 2025; M. Zhu et al. 2025a, b)DiabetesType 1 diabetes simulator; onset and progression modelingMechanistic glucose–insulin models with adaptive controlUVA/Padova simulators; clinical validation cohortsProspective/In silicoBenchmark for artificial pancreas testing; improved dosing algorithms(Cobelli et al. 2011; Visentin et al. 2018)NeurodegenerativeDisease modeling, cognitive decline monitoring, rehabMulti-scale neural models; wearables analyticsCohort studiesRetrospective/ProspectiveSupports personalized rehabilitation and early detection(Dominguez-Gortaire et al. 2025; Jirsa et al. 2017) Disease domains and conditions where in silico twins (ISTs) have been applied. Data flow and evidence types in legend. Representative references: (Al-Kabani et al. 2025); (Visentin et al. 2018); (Kim et al. 2020). Legend: data flow = clinical cohorts/registries → model building → validation and application. Update cadence: prospective registries = ongoing; literature-derived = retrospective Treatment simulation and optimization The application of ISTs for treatment simulation and optimization marks a significant advancement in personalized medicine, enabling clinicians to tailor interventions with unprecedented precision (Garg 2025; Panchpuri et al. 2025; Serrano et al. 2024). This section details how these digital models are used for personalized drug response modeling, real-time dosage optimization, and adaptive cancer therapy simulation. Personalized drug response modeling ISTs are instrumental in personalizing treatment by predicting how individual patients will respond to specific drugs based on their unique biological characteristics (Serrano et al. 2024). By integrating multimodal data, including genomics, proteomics, and metabolomics, AI models can identify biomarkers linked to drug efficacy and safety (Garg 2025; Koshechkin et al. 2022; Panchpuri et al. 2025; Serrano et al. 2024). This allows for the personalization of treatment regimens, such as dosing schedules, based on patient biomarkers and mutational status (Metzcar et al. 2025). ISTs can also explain how genetic backgrounds alter disease mechanisms, variability in disease manifestation, and drug responses, providing context-specific insights (Noutahi et al. 2025). This shift from generalized to personalized treatment is crucial for reducing adverse effects and improving overall patient outcomes (Alavi et al. 2024a, b, c, d; Serrano et al. 2024). Optimizing dosages in real-time The development of AI-driven predictive models allows for the real-time optimization of drug dosages, minimizing adverse effects and ensuring treatment adherence (Panchpuri et al. 2025). QSP models, for example, can create patient-specific ISTs to simulate T-cell kinetics in therapies for cancers, predicting the optimal dose for enhanced persistence and functional outcomes (Mirakhori & Niazi 2025). Model-informed precision dosing is an advanced application that leverages continuous data streams from wearables and other biosensors to adjust therapeutic interventions dynamically (Panchpuri et al. 2025). This real-time adaptability allows healthcare providers to fine-tune medication regimens based on immediate physiological responses, moving beyond static prescriptions to truly adaptive and individualized patient care (Garg 2025; Panchpuri et al. 2025). Adaptive cancer therapy simulation ISTs play a transformative role in designing and optimizing adaptive cancer therapies. These models can simulate outcomes for various therapeutic interventions, including assessing anti-PD-1, CAR-T, and bispecific antibodies within a human immune context (Al-Kabani et al. 2025; Garg 2025). They predict variability in treatment outcomes and guide model-driven combination dose selection using multi-objective synergy optimization (Metzcar et al. 2025). AI-driven platforms offer high scalability, facilitating hypothesis generation, patient stratification, and clinical trial design for cancer treatments (Al-Kabani et al. 2025; Mirakhori & Niazi 2025). For example, an IST can model disease progression under different chemotherapy or immunotherapy agents, enabling physicians to test multiple treatment pathways in silico before clinical application, reducing risks and improving efficiency (Garg 2025). Surgical and interventional planning ISTs are revolutionizing surgical and interventional planning by providing highly detailed, patient-specific virtual environments for complex procedures (Chen et al. 2024; Panchpuri et al. 2025). These digital models enable surgeons to rehearse and optimize interventions, significantly enhancing precision, reducing risks, and improving patient outcomes. ISTs are increasingly transforming surgical planning and interventional guidance through high-fidelity, patient-specific simulations that can predict procedural outcomes before the first incision. These platforms extend beyond visual anatomical modeling by incorporating biomechanical, hemodynamic, and electrophysiological simulations, enabling clinicians to test “virtual surgeries” and optimize procedural strategies. Regulatory-approved case studies demonstrate the translational readiness of this technology. The HeartFlow FFRCT platform (FDA-cleared, CE-marked) integrates coronary CT angiography with computational fluid dynamics (CFD) to estimate fractional flow reserve (FFR) non-invasively. Validated across over 2,500 patients in the DISCOVER-FLOW and NXT trials, it achieved diagnostic accuracies of 81–86% versus invasive FFR. It is now used in more than 100,000 patients annually as a standard adjunct to coronary diagnosis (Min et al. 2012; Nørgaard et al. 2014). The HeartFlow FFRCT platform (FDA-cleared, CE-marked) integrates coronary CT angiography with computational fluid dynamics (CFD) to estimate fractional flow reserve (FFR) non-invasively. Validated across over 2,500 patients in the DISCOVER-FLOW and NXT trials, it achieved diagnostic accuracies of 81–86% versus invasive FFR. It is now used in more than 100,000 patients annually as a standard adjunct to coronary diagnosis (Min et al. 2012; Nørgaard et al. 2014). Materialise Mimics (FDA-cleared, CE-marked) provides a certified platform for 3D reconstruction and surgical simulation across orthopedics, craniofacial, and cardiovascular domains. The system enables regulatory-compliant, personalized anatomical modeling from DICOM data, often serving as the geometric basis for downstream IST simulations (Giannopoulos et al. 2016; Kamio et al. 2018). Materialise Mimics (FDA-cleared, CE-marked) provides a certified platform for 3D reconstruction and surgical simulation across orthopedics, craniofacial, and cardiovascular domains. The system enables regulatory-compliant, personalized anatomical modeling from DICOM data, often serving as the geometric basis for downstream IST simulations (Giannopoulos et al. 2016; Kamio et al. 2018). The FEops HEARTguide (CE-marked) represents a fully in silico predictive platform for structural heart interventions, simulating device deployment (e.g., TAVI, LAA occluders) within patient-specific anatomies. In a prospective multicenter trial (n = 200), FEops reduced procedural complications by 28% compared to standard imaging-based planning (De Backer et al. 2023; Garot et al. 2020). The FEops HEARTguide (CE-marked) represents a fully in silico predictive platform for structural heart interventions, simulating device deployment (e.g., TAVI, LAA occluders) within patient-specific anatomies. In a prospective multicenter trial (n = 200), FEops reduced procedural complications by 28% compared to standard imaging-based planning (De Backer et al. 2023; Garot et al. 2020). Together, these examples demonstrate that regulatory-approved IST implementations are already redefining surgical planning—offering quantitative, personalized, and ethically validated digital simulations that improve safety and outcomes. Looking forward, emerging IST frameworks will integrate real-time feedback from intraoperative imaging and edge computing to enable adaptive, closed-loop guidance. As regulatory frameworks (FDA, EMA, SFDA) increasingly recognize model-informed surgical planning, the clinical role of ISTs is expected to expand rapidly across cardiovascular, orthopedic, and neurosurgical domains. Heart valve replacement In silico models and DTs are increasingly utilized in cardiovascular medicine to simulate complex physiological processes and plan intricate interventions, such as heart valve replacement. Computational fluid dynamics (CFD) modeling, for instance, can simulate blood flow and pressure dynamics in various anatomical configurations, allowing surgeons to virtually test different prosthetic valve designs and placements before actual surgery (Bailoor et al. 2021; Berg et al. 2019). This predictive capability helps in optimizing surgical outcomes by ensuring the best fit and function of the replacement valve, thereby minimizing complications and improving patient safety (Bailoor et al. 2021; Berg et al. 2019). While specific mentions of "heart valve replacement" ISTs are not detailed in the sources, the broader concept of computational modeling of hemodynamics in aneurysms (Bracamonte et al. 2022; Hamet & Tremblay 2017) and ISTs for cardiotoxicity (Passini et al. 2017) demonstrates the foundational capabilities that could be extended to valve procedures. The use of medical imaging with AI for diagnosis further supports the generation of highly accurate anatomical models for such planning. Virtual anatomical modeling for neurosurgery Virtual anatomical modeling using ISTs revolutionizes neurosurgical planning by creating highly detailed and patient-specific replicas of the brain and spinal cord. Digital computers are routinely used to generate three-dimensional models of target proteins (Braga & Rawal 2025) and intricate brain structures (Iglesias 2023). This allows neurosurgeons to simulate complex procedures, such as tumor resection or aneurysm clipping, in a virtual environment, minimizing risks and enhancing precision (Chen et al. 2024). The Virtual Brain project is a prominent example of creating an in silico representation of the brain for research and clinical applications (Sanz Leon et al. 2013). AI-powered analysis of magnetic resonance imaging (MRI) data further refines these models, enabling robust ML segmentation for large-scale analysis of heterogeneous clinical brain MRI datasets (Mennella et al. 2024). Furthermore, AI-based patient ISTs are being explored in rehabilitation and physical therapy, indicating a broader application of these models in neuro-recovery planning (Mikołajewska et al. 2024). Neuro-twin systems – validation and challenges: The Virtual Brain (TVB) has enabled patient-specific simulations of epileptic networks and lesion propagation, improving surgical targeting accuracy (Goldman et al. 2023). Its strength lies in biophysical realism and modular coupling to MRI, EEG, and LGE-MRI modalities. Nonetheless, model personalization remains computationally intensive, and attempts to predict seizure onset timing show variable reproducibility across cohorts (Proix et al. 2018). Overfitting risks emerge when optimizing large parameter spaces from limited neurophysiological data, motivating hybrid strategies combining transfer learning and population priors to constrain adaptive updates. Drug discovery and development ISTs are profoundly transforming drug discovery and development by enabling more efficient, cost-effective, and targeted approaches (An et al. 2011; Braga & Rawal 2025). These digital tools accelerate the identification of novel compounds, improve toxicity assessment, and significantly reduce preclinical failure rates. Virtual screening and compound testing ISTs are profoundly impacting drug discovery by enabling high-throughput virtual screening and compound testing (Chen et al. 2024; Koshechkin et al. 2022). This involves using computational methods and computer simulations to study and analyze complex systems, creating virtual representations of real-world systems (Malheiro et al. 2023). Generative AI models, such as generative adversarial networks (GANs) and variational autoencoders (VAEs), can design novel drug-like molecules from scratch and explore billions of potential compounds in silico within days (Braga & Rawal 2025; Garg 2025; Noutahi et al. 2025). This significantly accelerates the lead identification phase, which was traditionally the longest part of drug discovery (Garg 2025). AI also aids in predicting drug–protein binding affinities (RadhaMahendran et al. 2024), optimizing chemical reactions, and discovering new molecules (Malheiro et al. 2023). Platforms like Schrödinger Maestro (Patel et al. 2023), In silico Medicine (Brinkhaus et al. 2023), and Cyclica (Naveja-Romero 2022) leverage these capabilities for molecular modeling and simulations. Toxicity modeling ISTs play a critical role in predicting drug toxicity and adverse drug reactions early in the development pipeline (Chen et al. 2024). By simulating the behavior of virtual models, researchers can gain insights into underlying mechanisms and predict outcomes without costly physical experiments (Malheiro et al. 2023). This approach allows for the early toxicity assessment of potential drug candidates and the identification of potential failures of drug trials with respect to toxicity (Panchpuri et al. 2025). AI-driven models can analyze large datasets to predict off-target interactions for small molecules, a key aspect of safety profiling (Rao et al. 2019). The ability to assess toxicity in silico before preclinical in vivo experiments or clinical trials significantly enhances the safety profile of new drugs, offering a higher accuracy than traditional animal models in predicting clinical pro-arrhythmic cardiotoxicity (Metzcar et al. 2025). Role in reducing preclinical failure The integration of ISTs helps reduce preclinical failure rates by streamlining drug development processes (An 2001; An et al. 2011). Virtual clinical trials, based on extensions of pharmacokinetic/pharmacodynamic (PK/PD) modeling and systems biology, can incorporate preclinical (in vitro and in vivo) data to predict outcomes before patient trials (Metzcar et al. 2025). This makes trials more likely to succeed and reduces variability (Metzcar et al. 2025). By simulating dynamic and complex biological systems, in silico models elucidate how drug molecules interact with targets and behave within the human body, contributing to the design of safer and more effective therapeutics (Braga & Rawal 2025). This data-driven approach allows for the prioritization of promising compounds and minimizes the need for costly and time-consuming physical experiments, thereby mitigating risks and accelerating drug development (Braga & Rawal 2025; Malheiro et al. 2023; Panchpuri et al. 2025). Critical insights and current limitations: Across domains, IST applications illustrate both clear translational progress and persistent methodological gaps. Models validated in controlled settings—such as the UVA/Padova diabetes twin or HeartFlow FFR-CT—have demonstrated reproducible gains in clinical decision support. Conversely, challenges include false positives in rare disease prediction (due to data imbalance), drift and overfitting in continuously adaptive therapy models, and computational cost in high-dimensional parameter calibration. Addressing these issues will require transparent reporting of uncertainty, external validation datasets, and standardized V&V metrics as endorsed by regulatory frameworks (Viceconti et al. 2021). Case studies and real-world implementations The virtual physiological human (EU) The VPH represents a long-standing international initiative, initially launched in the late 1990 s as the Physiome project (Loewe et al. 2025; Prado-Velasco 2025). Its fundamental objective is to comprehensively describe and predict human physiology across various spatial and temporal scales, from proteins to organs (Loewe et al. 2025; Prado-Velasco 2025). Over time, the VPH concept evolved, shifting its primary focus from creating a singular, universal model of "the human" to developing disease-specific models and addressing inter-individual variability within patient cohorts, which can be leveraged for in silico clinical trials (Table 6) (Loewe et al. 2025). Table 6 has been updated to include success indicators summarizing validation outcomes and scalability metrics for key IST implementations. Table 6Comparative overview of representative in silico twin systems and frameworks across major clinical domains, extended with validation and scalability indicatorsDomain/SystemCore FocusModel TypeValidation Dataset/TrialSuccess Indicators/Clinical Validation StatusCardiology – HeartFlow FFR-CTNon-invasive coronary physiology modelingCFD-based physiological simulation with CT inputDISCOVER-FLOW & NXT trials (> 2,500 patients)FDA-cleared; diagnostic accuracy 81–86%; > 100 k patients used annuallyCardiology – FEops HEARTguideDevice–anatomy interaction (TAVI, LAAO)Finite-element structural modelingMulticenter prospective study (n = 200)CE-marked; reduced complications 28%; expanding to FDA submissionOncology – Physiomics Virtual TumourDrug response prediction/dose optimizationMechanistic PK/PD + Bayesian AI hybridPhase I–II oncology trialsIntegrated in > 15 studies; model-informed dose adjustments; commercial serviceMetabolism – UVA/Padova T1D TwinGlucose–insulin dynamicsDifferential-equation + adaptive controlSimulated + clinical validation (n≈300)Used in > 10 FDA-cleared AID algorithms; benchmark simulator standardNeurology – Dassault BIOVIA Living BrainNeurovascular and lesion modelingMulti-scale finite-element + MLInternal validation + pilot clinical useOngoing clinical pilots; part of FDA–industry collaboration consortiumOncology – Quris-AI BioTwinMicrophysiological + AI drug safety predictionLab-on-chip + generative MLPreclinical & collaborative pharma trialsValidated vs in vivo toxicity data; multiple pharma partners (Pfizer, Merck)Multi-organ/Regulatory Sandbox – FDA V&V40/EMA M&S FrameworkCross-domain model verification & validationMeta-framework (policy-driven)–Adopted by > 50 institutions; reference for regulatory qualification of ISTs Comparative overview of representative in silico twin systems and frameworks across major clinical domains, extended with validation and scalability indicators The VPH infrastructure aims to be a distributed, collaborative knowledge repository and simulation platform of quantitative human (patho)physiology (Viceconti et al. 2023). This initiative seeks to accelerate the development, integration, and adoption of patient-specific predictive computer models for clinical decision support and for refining, reducing, and replacing traditional in vitro, animal, and human experimentation in medical product assessment (Viceconti et al. 2023). It addresses the challenge of integrating complex and diverse data and models, emphasizing the need for open tools, data, and metadata standards, alongside strong community interactions (Loewe et al. 2025). The European Commission has actively supported this vision, even drafting a roadmap for the "Human IST" initiative (Dos Santos et al. 2024; Lesage et al. 2023). TwinHeart (cardiac digital twin) The concept of a cardiac IST involves creating a personalized computational model of an individual patient's heart, mimicking its structure and function (Loewe et al. 2025). These models are developed using anatomical and functional personalization, often derived from medical imaging data like computed tomography (CT) or MRI for anatomical twinning, and electrocardiograms (ECGs) for functional twinning (Loewe et al. 2025). Cardiac ISTs are employed for individual risk prediction, clinical decision support, and therapy planning, enabling in silico evaluation of interventions before application to a patient, aligning with personalized medicine (Corral-Acero et al. 2020; Loewe et al. 2025). Prakosa et al. (Prakosa et al. 2018) developed the Virtual-heart Ablation-target (VAAT) technology to guide ablation of infarct-related ventricular tachycardia (Prakosa et al. 2018). This personalized approach, built from cardiac imaging and computational modeling, non-invasively identifies optimal ablation targets, potentially reducing the need for repeated ablations (Prakosa et al. 2018). Their retrospective animal and human studies, along with a prospective feasibility study, demonstrated that VAAT-guided ablations were feasible within clinical workflows (Prakosa et al. 2018). Lumens et al. (Lumens et al. 2015) utilized computer simulations to distinguish patterns of left ventricular mechanical discoordination, differentiating those caused by electromechanical substrates (responsive to Cardiac Resynchronization Therapy (CRT)) from those linked to non-electrical factors like regional hypocontractility or scar tissue (unresponsive to CRT) (Lumens et al. 2015). This mechanistic understanding led to the development of a novel radial strain-based metric, the SSI (Strain Synchrony Index), which proved effective in identifying patients who benefited more favorably from CRT, particularly those with intermediate ECG criteria where CRT response is often uncertain (Lumens et al. 2015). Furthermore, Martinez-Velazquez et al. (Martinez-Velazquez et al. 2019) presented "Cardio Twin," a DT architecture for Ischemic Heart Disease (IHD) detection designed to run efficiently on edge devices (Martinez-Velazquez et al. 2019). This system leverages a Convolutional Neural Network (CNN), trained on ECG data, to classify non-myocardial and myocardial conditions, achieving an 85.77% accuracy with a rapid classification time of 4.8 s per sample. Cardio Twin is built to integrate data from various sensors and aims to provide real-time monitoring and decision support, even facilitating direct communication with smart services for emergency situations, demonstrating the viability of its AI-Inference engine on edge devices for predictive and preventive applications (Martinez-Velazquez et al. 2019). Siemens healthineers and GE healthcare initiatives IST technology is being actively adopted by major healthcare companies to optimize operational efficiency and improve clinical decision-making. GE Healthcare, for instance, has implemented its "Capacity Command Center" at Johns Hopkins Hospital in Baltimore (Khan et al. 2022b; Khan et al. 2022c). This initiative utilizes IST concepts, predictive analytics platforms, and advanced AI capabilities to transform vast amounts of patient data into actionable intelligence, thereby optimizing hospital management and improving decision-making processes (Khan et al. 2022b; Khan et al. 2022c). Similarly, Siemens Healthineers has played a crucial role in enhancing the efficiency of Mater Private Hospital (MPH) in Dublin through the deployment of IST technology (Khan et al. 2022b; Khan et al. 2022c). Their efforts focused particularly on the radiology department, where AI computer models were utilized to optimize departmental operations. Through these IST implementations, MPH has successfully addressed challenges such as increasing clinical complexity, aging infrastructure, delays, growing patient demands, and the processing of large volumes of data (Khan et al. 2022b; Khan et al. 2022c). These examples underscore how IST technologies are being applied beyond direct patient care to optimize healthcare systems, providing benefits in terms of cost reduction and improved service delivery (Khan et al. 2022b). Predictive twins in ICU settings In the complex and time-sensitive environment of Intensive Care Units (ICUs), ISTs offer significant potential for processing massive amounts of data and supporting critical decisions (Eslami et al. 2024; Hennigs et al. 2024). These predictive twins aim to fuse sensor data with physiological knowledge to create an individualized digital representation of each patient, which can be used for diagnosis and therapy (Hennigs et al. 2024). Eslami et al. (Eslami et al. 2024) demonstrated the feasibility of developing ICU specialty-adapted ISTs using Large Language Models (LLMs), specifically LLaMA-3, fine-tuned with Low-Rank Adapters (LoRA), for medication prediction (Eslami et al. 2024). Their research revealed that specialty-specific training on notes from a Medical ICU significantly enhanced medication recommendation accuracy compared to generalized or zero-shot approaches. This approach reflects how ICU specialists develop expertise through practice within their specific critical care environments and highlights the ISTs' potential to alleviate cognitive burden and reduce medication errors by adapting to nuanced practice patterns and health system guidelines (Eslami et al. 2024). In parallel, Hennigs et al. (Hennigs et al. 2024) introduced a IST-based monitoring tool specifically for mechanically ventilated patients in the ICU (Hennigs et al. 2024). This tool employs a sensor fusion approach, integrating data from various modalities such as surface electromyography (sEMG) and electrical impedance tomography (EIT), to continuously estimate physiological parameters. The tool's front-end provides clinicians with a condensed overview of a patient's past and current status. A notable "prediction tab" allows clinicians to simulate changes in ventilator settings (e.g., driving pressure, respiratory rate, inspiratory oxygen fraction) and observe the predicted physiological effects, forming a crucial basis for a medical DT in respiratory therapy (Hennigs et al. 2024). Model development and validation The advancement of ISTs in precision medicine hinges critically on robust model development and thorough validation. This encompasses not only the initial construction of sophisticated computational models but also their continuous refinement, personalization, and rigorous evaluation against real-world data and clinical outcomes. The goal is to ensure that these digital representations accurately reflect individual patient physiology and disease progression, providing reliable tools for AI-driven diagnosis, treatment simulation, and drug discovery. Model calibration and personalization Model calibration and personalization are fundamental steps in creating accurate ISTs, ensuring that generic models are adapted to reflect individual patient characteristics and responses (Abdollahi et al. 2024; Xu & Wang 2025). The development of Theranostic Digital Twins (TDTs), for instance, begins with comprehensive patient data acquisition, encompassing multimodal and multi-scale information (Abdollahi et al. 2024). This includes diverse inputs such as blood tests for specific biomarkers (e.g., tumor markers), urine tests for metabolic indicators, and advanced imaging modalities like PET and CT scans for anatomical and functional insights (Abdollahi et al. 2024). Furthermore, wearable body-attached sensors provide continuous data on vital signs, heart rate, and activity levels, ensuring a holistic representation of the patient's health status. Physicochemical data, including a radiopharmaceutical's physical half-life, dose rate, and emission type, are also integral for TDT models, as these properties significantly influence therapeutic delivery, effectiveness, and toxicity (Abdollahi et al. 2024). A critical aspect of personalization is parameter estimation, an inverse problem that tunes the model to specific patient data (Abdollahi et al. 2024). This can be achieved through conventional mathematical methods like linear and non-linear least-square fitting or genetic algorithms, as well as modern AI approaches such as Biology-Informed Neural Networks (BINNs). BINNs are particularly promising due to their potential for interpretability and explainability, having demonstrated utility in stratifying prostate cancer patients (Abdollahi et al. 2024). In musculoskeletal modeling, generic designs are often insufficient; personalization is crucial to accurately represent individual muscle properties, which do not scale linearly with bones (Xu & Wang 2025). For example, studies integrate patient-specific Physiological Cross-Sectional Area (PCSA) values and electromyography (EMG) data to precisely reproduce physiological tests like Maximum Voluntary Isometric Contraction (MVIC), significantly improving prediction accuracy (Bottin et al. 2024). Similarly, ISTs for breast cancer management use longitudinal MRI data to predict pathological complete response (pCR), allowing for individualized treatment escalation or non-escalation strategies based on predicted outcomes (Wu et al. 2025). These models can operate in a "silent mode" during observational trials, predicting responses to actual and standard of care (SOC) treatments without influencing clinical decisions. Ultimately, models like the Reinforcement Learning-based Dynamic Insulin Titration (RL-DITR) system, described in a study using Reinforcement Learning (RL), optimize treatment regimens by learning from patient response data through iterative interactions between a policy model and a patient model (Wang et al. 2023). For robust in silico models, models run synchronously with physical entities and are continually refined through calibration strategies, supporting accurate estimation, optimization, and prediction (Liu et al. 2019). Validation and verification Validation involves rigorously evaluating a model's performance on new data or in real-world settings to ensure its accuracy, reliability, and effectiveness (Abdollahi et al. 2024). For ISTs, this necessitates thorough validation and calibration of predictions using data from patient cohorts to ensure reliability and accuracy in real-world clinical settings (Fig. 4A) (Abdollahi et al. 2024). Verification, on the other hand, ensures that the model's predictions accurately replicate actual patient data and maintain the integrity of cross-correlations between features (Ferle et al. 2025). The Verification, Validation, and Uncertainty Quantification (VV&UQ) framework, supported by the National Research Council's report on assessing complex models, emphasizes these steps (Ferle et al. 2025; Lewalle et al. 2025). A (A) End-to-end workflow for model development and validation, progressing from raw patient data to calibrated digital twin (DT) construction and external validation. Arrows indicate data flow from retrospective clinical datasets to model calibration, followed by prospective simulation and comparison with observed outcomes. (B) External validation of the models using the German-Speaking Myeloma Multicenter Group (GMMG/MM5) study dataset. (i) Aggregated mean squared error (MSE) of the forecasting model compared with baseline methods—last observation carried forward (LOCF) and moving average (MA)—across forecasting horizons of 3, 6, 9, 12, and 15 months. Cross and asterisk symbols denote p < 0.1 and p < 0.05, respectively, for comparisons between the forecasting model and LOCF; dagger and double dagger symbols represent the same significance thresholds for comparisons between the forecasting model and MA. (ii) Correlation coefficients from actual data juxtaposed with those from forecasted data across forecasting horizons, with dashed lines representing the best-fit lines and dotted lines showing the expected lines of best fit for m = 1 and b = 0. All parameters are reported as mean ± standard deviation across cross-validation folds. (iii–iv) Discriminative performance of the combined model for predicting progression events, expressed as area under the receiver operating characteristic curve (AUROC) (iii) and area under the precision–recall curve (AUPRC) (iv) across forecasting horizons, with bars and error bars indicating mean ± standard deviation across folds. (v–vi) Baseline performance of the annotation model: receiver operating characteristic (ROC) curve (v) with AUROC = 0.87 ± 0.01, and precision–recall curve (PRC) (vi) with AUPRC = 0.27 ± 0.01, substantially exceeding the chance level of 0.03 ± 0.00. (vii–viii) Heatmaps depicting the combined model’s discriminative performance (AUROC in vii, AUPRC in viii) as a function of the prior observation period and forecasting horizon. Figure reproduced with permission from Ferle et al. (Ferle et al. 2025), Copyright 2025 Nature Publishing Group A critical challenge in this domain is the lack of external validation on independent data from other institutions, which can undermine AI applications and limit generalizability (Xu & Wang 2025). However, a study on predicting multiple myeloma progression demonstrated robust external validation using an independent dataset (GMMG-MM5 study dataset), confirming the model's resilience to differences in data collection and demographic variations (Fig. 4B) (Ferle et al. 2025). This study also employed a fivefold cross-validation approach to quantify variability and assess generalizability and robustness. Quantitative metrics such as AUROC (Area Under the Receiver Operating Characteristic curve) and AUPRC (Area Under the Precision-Recall curve) are crucial for assessing discrimination and predictiveness (Ferle et al. 2025). For regression models, RMSE (Root Mean Square Error), MAE (Mean Absolute Error), and R2 (Coefficient of Determination) are utilized (Imans et al. 2024). Uncertainty Quantification (UQ) is addressed by evaluating the impact of varying amounts of historical patient data and different forecasting horizons on model performance metrics (Ferle et al. 2025). The concept of in silico clinical trials is gaining significant traction, with the FDA increasingly providing guidance on their use in submissions (Abdollahi et al. 2024; Xu & Wang 2025). New FDA legislation in 2023 specifically mandates that new drugs can proceed to human trials following successful "non-clinical tests," including "Computer modeling," as an alternative to animal testing (Abdollahi et al. 2024; Xu & Wang 2025). This shift underscores the growing trust in computational models to predict drug responses and optimize therapeutic strategies. For example, a RL-DITR system underwent a comprehensive evaluation process for clinical integration, including multi-center retrospective studies and prospective studies (Wang et al. 2023). Retrospective analyses compared the AI model's performance against human physicians in insulin dosage titration, showing performance in terms of predicted error, dosage adjustment agreement, effectiveness, safety, and overall acceptability across different physician experience levels (Wang et al. 2023). A prospective study further assessed the effectiveness, safety, overall acceptability, and adoption rate of AI-generated regimens directly at the bedside (Wang et al. 2023). Verification, validation, and uncertainty quantification (V&V & UQ) Robust V&V and UQ are essential for clinical translation of ISTs (Table 7). A pragmatic pipeline comprising the following components is proposed: Model verification — Ensure code correctness and numerical fidelity through unit and integration testing, solver convergence studies, and version-controlled repositories with continuous integration. Follow ASME V&V 40 principles for mechanistic/model credibility where applicable. Model verification — Ensure code correctness and numerical fidelity through unit and integration testing, solver convergence studies, and version-controlled repositories with continuous integration. Follow ASME V&V 40 principles for mechanistic/model credibility where applicable. Calibration strategies — Use hierarchical or Bayesian parameter estimation to personalize mechanistic model parameters when individual-level data are available. Report prior distributions, hyperparameters, and convergence diagnostics. Include posterior predictive checks and calibration plots (e.g., calibration slope and intercept). Calibration strategies — Use hierarchical or Bayesian parameter estimation to personalize mechanistic model parameters when individual-level data are available. Report prior distributions, hyperparameters, and convergence diagnostics. Include posterior predictive checks and calibration plots (e.g., calibration slope and intercept). Internal validation — Use cross-validation, temporal splits, and bootstrapping to estimate model optimism and uncertainty. Report performance with uncertainty intervals. Internal validation — Use cross-validation, temporal splits, and bootstrapping to estimate model optimism and uncertainty. Report performance with uncertainty intervals. External validation — Evaluate IST performance on geographically, temporally, or demographically distinct cohorts with a pre-specified analysis plan. Quantify performance degradation (ΔAUROC, calibration slope) and report results transparently following TRIPOD-ML guidance. External validation — Evaluate IST performance on geographically, temporally, or demographically distinct cohorts with a pre-specified analysis plan. Quantify performance degradation (ΔAUROC, calibration slope) and report results transparently following TRIPOD-ML guidance. Prospective pilots and clinical feasibility — Conduct IRB-approved prospective pilots or feasibility studies, report operational metrics (time-to-decision, clinician adherence) and clinical impact where feasible, and consider pre-registration to improve transparency (CONSORT-AI, SPIRIT-AI). Prospective pilots and clinical feasibility — Conduct IRB-approved prospective pilots or feasibility studies, report operational metrics (time-to-decision, clinician adherence) and clinical impact where feasible, and consider pre-registration to improve transparency (CONSORT-AI, SPIRIT-AI). Uncertainty quantification (UQ) — Propagate parameter uncertainty to predictions using Bayesian credible intervals, bootstrapped confidence intervals, and ensemble methods. Present 95% credible/confidence intervals and assess calibration of predictive distributions. Uncertainty quantification (UQ) — Propagate parameter uncertainty to predictions using Bayesian credible intervals, bootstrapped confidence intervals, and ensemble methods. Present 95% credible/confidence intervals and assess calibration of predictive distributions. Decision-focused evaluation — Evaluate clinical utility using decision-curve analysis, net benefit, and clinically meaningful metrics such as net reclassification improvement (NRI) for risk stratification models. Specify clinically relevant thresholds and report net benefit across those ranges. Decision-focused evaluation — Evaluate clinical utility using decision-curve analysis, net benefit, and clinically meaningful metrics such as net reclassification improvement (NRI) for risk stratification models. Specify clinically relevant thresholds and report net benefit across those ranges. Drift monitoring & maintenance — Implement continuous monitoring for data and concept drift with predefined thresholds that trigger model review or retraining. Maintain change logs, versioning, and post-deployment surveillance consistent with Good Machine Learning Practice and regulatory expectations. Drift monitoring & maintenance — Implement continuous monitoring for data and concept drift with predefined thresholds that trigger model review or retraining. Maintain change logs, versioning, and post-deployment surveillance consistent with Good Machine Learning Practice and regulatory expectations. Explainability & reporting — Provide interpretability analyses (feature importance stability, counterfactual checks), complete documentation of datasets and preprocessing, and open-source code when possible to enable reproducibility. Follow TRIPOD-ML and PROBAST-AI reporting recommendations. Explainability & reporting — Provide interpretability analyses (feature importance stability, counterfactual checks), complete documentation of datasets and preprocessing, and open-source code when possible to enable reproducibility. Follow TRIPOD-ML and PROBAST-AI reporting recommendations. Table 7Minimal IST V&V & UQ Checklist. A pragmatic checklist of verification, validation, uncertainty quantification, and regulatory documentation recommended for in silico twins (ISTs). For clinical deployment, expand per disease-specific and regulator-specific requirementsChecklist ItemSuggested Evidence/ProceduresSuggested Metrics/Acceptance CriteriaRelevant Guidance/ReportingModel Verification (code correctness, numerical solvers)Unit tests, integration tests, solver-convergence reports, version-controlled code repository, continuous integrationAll tests passing; numerical error < domain-specific tolerance (report)ASME V&V 40; general software engineering best practicesCalibration StrategyCalibration on training data with clear protocol; use of hierarchical/Bayesian parameter estimation where appropriate; report priors and hyperparametersConvergence diagnostics; posterior predictive checks; calibration curvesTRIPOD-ML; PROBAST-AIInternal ValidationCross-validation, bootstrapping, temporal/hospital split validation; report uncertainty intervalsStable performance across folds; optimistic bias assessed and reportedTRIPOD-MLExternal ValidationEvaluation on independent cohorts (geographically and temporally distinct), with pre-specified analysis planPerformance degradation quantified; target performance defined a priori (e.g., ΔAUROC, calibration slope)TRIPOD-ML; CONSORT-AI recommendations for external validationProspective Pilot/Clinical FeasibilityPre-registered pilot studies, IRB approval, real-time deployment logs, clinician feedbackOperational metrics: time-to-decision, clinician adherence; clinical impact endpoints where possibleCONSORT-AI; SPIRIT-AIUncertainty Quantification (UQ)Bootstrapping, Bayesian posterior intervals, ensemble approaches, probabilistic sensitivity analysis, propagation of parameter uncertainty to outputsReport 95% credible/confidence intervals; calibration of predictive distributionsPROBAST-AI guidance; statistical standardsDecision-focused EvaluationDecision curve analysis, net reclassification improvement (NRI), net benefit, cost-effectiveness where possiblePositive net benefit across clinically relevant threshold ranges; clinically meaningful NRIClinical decision analysis literature; CONSORT-AI for clinical trialsDrift Monitoring & MaintenanceData drift detectors, concept drift detection, routine re-calibration schedule, monitoring dashboards and alertingPre-specified drift thresholds triggering retraining or reviewGood Machine Learning Practice (GMLP); regulatory post-market surveillance guidanceExplainability & ReportingModel interpretability reports, feature importance stability, saliency when applicable; full reporting of datasets, code, and pre-processingTransparent documentation; reproducibility checklist completedTRIPOD-ML; PROBAST-AI; ASME V&V 40Regulatory & Ethical DocumentationRisk assessment, human factors analysis, data governance, privacy impact assessment, versioning and change logsRisk mitigation plans; documented consenting procedures and data governanceFDA/EMA guidance, ASME V&V 40 Minimal IST V&V & UQ Checklist. A pragmatic checklist of verification, validation, uncertainty quantification, and regulatory documentation recommended for in silico twins (ISTs). For clinical deployment, expand per disease-specific and regulator-specific requirements Fairness, bias, and health equity in in silico twin deployment As ISTs become embedded in clinical workflows, the risks of algorithmic bias and inequitable performance must be explicitly managed. Dataset shift—differences between development and deployment populations—can cause systematic degradation in prediction accuracy across demographic or clinical subgroups (e.g., ancestry, sex, comorbidity, socioeconomic status). Such disparities can propagate through feedback loops, reinforcing underdiagnosis or over-treatment in marginalized groups (Obermeyer et al. 2019). Subgroup performance analysis should be a mandatory component of model validation. Disaggregated reporting of key performance metrics (AUROC, calibration slope, Brier score, and decision-curve net benefit) across protected and clinically relevant strata provides transparency on where the model underperforms. Frameworks such as TRIPOD-ML, PROBAST-AI, and SPIRIT-AI have begun recommending this reporting as part of model disclosure standards (Collins et al. 2015; Group TC-AaS-AS 2019). Mitigation strategies include: Re-weighting and stratified sampling to balance subgroup representation in training data. Re-weighting and stratified sampling to balance subgroup representation in training data. Domain adaptation and transfer learning to adjust for site or population shifts without full retraining. Domain adaptation and transfer learning to adjust for site or population shifts without full retraining. Fairness-aware learning objectives (e.g., equalized odds, demographic parity penalties) integrated into model optimization. Fairness-aware learning objectives (e.g., equalized odds, demographic parity penalties) integrated into model optimization. Post-hoc calibration by subgroup or adaptive recalibration during deployment. Post-hoc calibration by subgroup or adaptive recalibration during deployment. Beyond algorithmic fairness, equitable deployment requires addressing barriers to access: unequal data collection infrastructure, language interfaces, and digital literacy. Governance frameworks should ensure diverse patient representation in FL nodes and equitable consent and benefit-sharing mechanisms. Operational reporting should include subgroup-specific confidence intervals, drift monitoring stratified by demographic axes, and fairness dashboards during post-market surveillance (Chen et al. 2018; Rajkomar et al. 2018). Such reporting aligns with the FDA Good Machine Learning Practice (GMLP) and EU AI Act requirements for transparency and human oversight. In summary, fairness in ISTs is not solely a data-science issue but a sociotechnical imperative—requiring co-design with diverse stakeholders, continuous bias audits, and transparent reporting of subgroup performance to ensure equitable clinical benefit. Regulatory pathway and real-world deployment Translating ISTs into clinical practice requires alignment with medical-device and drug-regulatory frameworks, robust change-control for adaptive systems, and sustained post-market evidence generation. Below we map typical IST use-cases to likely regulatory classifications and summarize key regulatory expectations and practical steps for real-world deployment. Regulatory classification by use-case Clinical decision support (CDS) functions that provide recommendations to clinicians (but leave the final decision to the clinician) may be excluded from the device definition in some cases; however, CDS that interpret or process patient data to drive diagnosis or therapy decisions will commonly be regulated as Software as a Medical Device (SaMD) and require premarket regulatory review (Administration. 2018, 2022). Clinical decision support (CDS) functions that provide recommendations to clinicians (but leave the final decision to the clinician) may be excluded from the device definition in some cases; however, CDS that interpret or process patient data to drive diagnosis or therapy decisions will commonly be regulated as Software as a Medical Device (SaMD) and require premarket regulatory review (Administration. 2018, 2022). Device-led digital twins that focus on hardware performance (e.g., ventilator digital twins for maintenance) and are tightly coupled to a device are often reviewed as part of the device lifecycle (device + software). If the model informs therapy delivery (actuators, closed-loop control), regulatory scrutiny increases commensurate with risk (e.g., 510(k)/De Novo or PMA pathways in the US) (Administration. 2018). Device-led digital twins that focus on hardware performance (e.g., ventilator digital twins for maintenance) and are tightly coupled to a device are often reviewed as part of the device lifecycle (device + software). If the model informs therapy delivery (actuators, closed-loop control), regulatory scrutiny increases commensurate with risk (e.g., 510(k)/De Novo or PMA pathways in the US) (Administration. 2018). Model-informed precision dosing (MIPD) and other pharmacometrics-driven ISTs that influence individualized medication dosing often fall within the intersection of drug and device regulatory pathways: models used to support dosing decisions for marketed drugs may be considered medical device software (SaMD) or supportive evidence in drug labeling if developed during drug development. The FDA and EMA have active programs and draft/final guidance on model-informed drug development (MIDD) and mechanistic model reporting that are relevant to MIPD deployment (Administration. 2024; Agency 2025). Model-informed precision dosing (MIPD) and other pharmacometrics-driven ISTs that influence individualized medication dosing often fall within the intersection of drug and device regulatory pathways: models used to support dosing decisions for marketed drugs may be considered medical device software (SaMD) or supportive evidence in drug labeling if developed during drug development. The FDA and EMA have active programs and draft/final guidance on model-informed drug development (MIDD) and mechanistic model reporting that are relevant to MIPD deployment (Administration. 2024; Agency 2025). Change control for learning systems Adaptive ISTs that update parameters or models post-deployment pose a unique regulatory challenge. Regulatory bodies have proposed mechanisms to permit controlled learning in the field, notably the FDA’s Predetermined Change Control Plan (PCCP)/guiding principles and related AI/ML device guidance for pre-specifying allowable modifications and performance monitoring strategies. Inclusion of a PCCP-style plan (description of modification types, associated risk assessments, and monitoring triggers) in the premarket submission or technical documentation supports iterative improvement while preserving a reasonable assurance of safety and effectiveness. We recommend that ISTs adopt a predetermined change control plan that clearly defines permissible update classes, validation requirements for each class, rollback procedures, and clinical governance (Administration. 2025a, 2025b). Post-market surveillance and real-world evidence Post-market activities should include continuous data- and concept-drift monitoring, clinician-facing performance dashboards, scheduled re-calibration routines, and pre-specified thresholds that trigger re-validation or regulatory notification. Collecting real-world evidence (RWE) through prospective pilots, registries, or pragmatic trials provides the necessary outcomes data to demonstrate clinical utility and inform updates to model priors and decision thresholds. Where ISTs inform dosing or critical therapy decisions, prospective clinical pilots with pre-specified endpoints and safety monitoring (aligned with CONSORT-AI/SPIRIT-AI principles) are strongly recommended (Administration. 2022, 2024). Regulatory resources and practical steps We advise authors and developers to: (1) engage early with regulators via MIDD/paired meeting programs or pre-submission meetings to clarify evidentiary expectations; (2) create and document a PCCP-style change control plan for any adaptive components; (3) implement a minimal IST V&V checklist (Table 7) including prospective pilot plans and RWE collection; and (4) ensure transparent reporting per TRIPOD-ML/PROBAST-AI and device reporting where appropriate. The FDA has explicit resources on MIDD, SaMD/CDS, and AI/ML device considerations; the EMA is formalizing guidance on mechanistic model assessment and reporting—developers should reference these evolving guidances during planning and submission (Administration. 2022, 2024; Agency 2025). Example mappings (short) An IST that provides real-time, patient-specific dosing recommendations (MIPD) to an ICU clinician would likely be considered SaMD with high-risk implications and should be accompanied by prospective feasibility/pilot data, detailed PCCP documentation if adaptive, and post-market performance monitoring (Administration. 2024). An IST used solely for sponsor-led model exploration during drug development (supporting labeling) would typically follow MIDD engagement processes and may be submitted as part of the drug dossier rather than as separate device software (Administration. 2024; Agency 2025). Commercial and regulatory adoption of in silico twins The IST paradigm is transitioning from academic proof-of-concept models to commercially deployed and regulatory-recognized frameworks across the pharmaceutical, medical-device, and digital-health industries. In the biopharmaceutical domain, ISTs are being incorporated into model-informed drug development (MIDD) pipelines to accelerate trial design, dose optimization, and patient stratification (Morrison et al. 2019; Viceconti et al. 2021). Industry adoption is accelerating. Platforms such as Dassault Systèmes’ BIOVIA Living Heart and Living Brain projects—developed in collaboration with the U.S. FDA’s Digital Twin Consortium—provide certified simulation environments for cardiovascular and neurological device testing (Viceconti et al. 2021). In silico Medicine has integrated AI-driven generative chemistry with virtual twin modeling to design and validate drug candidates in silico (Ivanenkov et al. 2023). Physiomics PLC’s Virtual Tumour model and Unlearn. AI’s synthetic control arms are being actively deployed in oncology and clinical-trial optimization, with the latter receiving FDA recognition in Type C meetings for hybrid trials combining virtual and real participants (Elvatun et al. 2025; Pasculli et al. 2025; Popat et al. 2022). Startups such as Quris-AI are extending these paradigms to microphysiological systems linked to computational twins, advancing predictive preclinical validation. Regulatory engagement has also become increasingly formalized. The U.S. FDA’s Model-Informed Drug Development (MIDD) Framework and its 2023 Digital Health Technologies for Clinical Investigations Guidance define pathways for the verification, validation, and regulatory submission of quantitative models (Force et al. 2025; Madabushi et al. 2022; Mitra et al. 2024). Similarly, the European Medicines Agency (EMA) has issued a Reflection Paper on the Qualification of Modelling and Simulation for Drug Development and Regulatory Assessment, outlining procedures for model qualification and evidence integration (Musuamba et al. 2021; Walenga et al. 2025). National agencies—including the Saudi FDA (SFDA)—are aligning with these global policies by promoting AI/ML transparency, continuous learning systems, and data-provenance standards consistent with international digital health guidance. Collectively, these developments demonstrate a growing convergence between industry innovation and regulatory acceptance. As standardized V&V frameworks mature (Viceconti et al. 2021), ISTs are poised to underpin adaptive regulatory submissions, hybrid clinical trials, and continuous-learning healthcare systems that integrate real-world data with predictive simulation. Data governance, privacy, and federated learning Successful deployment of ISTs depends on robust data governance that balances utility with patient privacy and agency (Table 8). Below we outline pragmatic elements of a governance strategy for ISTs and describe privacy-enhancing technologies and cross-site harmonization practices. Table 8IST Data governance and stakeholder responsibilitiesStakeholderPrimary ResponsibilitiesData & Privacy Controls/ExamplesPatient/ParticipantConsent decisions; data subject rights (access, rectification, withdrawal); preferences for data sharing; receive summaries of model-derived decisionsGranular consent options; consent refresh for major model changes; access logs; patient-facing dashboardsHealthcare Provider/ClinicianProvide clinical data; interpret IST outputs; escalate safety concerns; enforce clinical governanceRole-based access; audit trails; explainability summaries for clinical decisionsHospital/Health System (Data Controller)Data stewardship; ensure lawful basis for processing; enforce data minimization and retention policies; oversee deployment & monitoringData minimization; de-identification; secure enclaves; data use agreements (DUAs); provenance logsModel Developer/Vendor (Data Processor)Develop models; perform training/validation; implement change-control, versioning, and security; provide transparency documentationSecure MPC/Homomorphic Encryption for training; differential privacy for aggregated outputs; code repositories and CI/CD with audit logsRegulator/HTA/PayerSet safety and efficacy standards; post-market surveillance requirements; reimbursement/coverage decisionsRequire documented V&V, PCCP, RWE collection plans; compliance checksEthics/IRB/Governance BoardApprove study protocols and deployment pilots; monitor ethical issues and bias; oversee community engagementEthics approval records; community consultation documentation; bias auditsThird-party Auditor/Independent ReviewerConduct independent audits of data governance, security, and model performance; verify compliance with reporting standardsAudit reports; attestation letters; penetration testing reportsData Infrastructure/Cloud ProviderProvide secure compute and storage; implement encryption at rest/in transit; support logging and provenanceHSMs, encryption, IAM, secure enclaves, SLAsLegal/Privacy OfficerEnsure contractual terms, DUAs, DPIAs (data protection impact assessments), and compliance with GDPR/other privacy lawsDPIA documents; contractual templates; breach response plans IST Data governance and stakeholder responsibilities Data minimization and provenance. Adopt the principle of data minimization: collect only data elements necessary for the intended IST function, and store derived features rather than raw identifiers when possible. Maintain detailed provenance metadata (source, timestamp, transformation, version) for every data element and model artifact to enable reproducibility and regulatory audit. Document data retention schedules and deletion procedures in Data Use Agreements (DUAs) and privacy impact assessments (DPIAs). Privacy-enhancing technologies (PETs). Consider PETs appropriate to the threat model and utility requirements: Federated learning (FL) enables training across multiple sites without centralizing raw data; aggregation can be combined with secure aggregation protocols to reduce leakage. Federated learning (FL) enables training across multiple sites without centralizing raw data; aggregation can be combined with secure aggregation protocols to reduce leakage. Secure multi-party computation (MPC) allows joint computation without revealing raw inputs to other parties, suitable for collaborative training with strict privacy constraints. Secure multi-party computation (MPC) allows joint computation without revealing raw inputs to other parties, suitable for collaborative training with strict privacy constraints. Homomorphic encryption (HE) enables computation on encrypted data but currently carries high computational cost; it may be appropriate for targeted cryptographic operations (e.g., encrypted inference in the cloud). Homomorphic encryption (HE) enables computation on encrypted data but currently carries high computational cost; it may be appropriate for targeted cryptographic operations (e.g., encrypted inference in the cloud). Differential privacy (DP) offers formal privacy guarantees for published outputs or model updates; it should be parameterized to balance privacy and model utility. Differential privacy (DP) offers formal privacy guarantees for published outputs or model updates; it should be parameterized to balance privacy and model utility. Cross-site schema harmonization and provenance. Harmonize data across sites using a common data model (e.g., OMOP or FHIR mappings) and shared terminology services. Implement rigorous ETL pipelines with version-controlled mapping scripts, automated validation checks, and a provenance ledger documenting transformations. These steps reduce semantic drift and improve external validity. Governance models, ownership, and patient agency. Define roles clearly: the health system typically acts as data controller, the model developer as processor, and patients retain data subject rights under applicable law. Governance agreements should clarify ownership of derivative artifacts (model weights, synthetic datasets), licensing, and permitted commercial use. Implement patient-facing controls: granular consent choices, concise model-use summaries, mechanisms for consent refresh when models change materially, and accessible summaries of IST-driven decisions. Operational safeguards and accountability. Include technical safeguards (role-based access control, encryption at rest/in transit, secure enclaves) and organizational safeguards (IRB/ethics oversight, independent audits, breach response plans). Record and report provenance and audit logs to facilitate incident investigation and regulatory inspections. Operational constraints and system-level considerations The practical deployment of ISTs in clinical environments requires explicit attention to operational performance, reliability, and sustainability constraints (Table 9). While most existing literature focuses on algorithmic innovation, successful bedside implementation depends on defined service-level agreements (SLAs) between model developers, hospital IT departments, and clinical end-users. Table 9Representative service-level agreements (SLAs) targets and operational benchmarks for clinical deployment of in silico system (IST)ConstraintRecommended Target/BenchmarkRationale/NotesEnd-to-end latency (bedside inference) ≤ 2 s (edge)/≤ 10 s (cloud-assisted)Enables real-time clinical responsiveness in high-acuity workflowsSystem uptime/availability ≥ 99.9% annual uptimeMeets clinical-grade reliability standards and safety–critical requirementsOffline operation (edge resilience) ≥ 24 h autonomous operationEnsures continuity during network or power disruptions; local buffering of data and inferenceCost efficiency < $0.05 per inference (approx. 100 ms GPU compute)Illustrative target aligned with sustainable hospital IT budgetsEnergy efficiency/carbon footprint ≤ 0.5 Wh per inference; CO₂ reporting per training epochSupports 'green AI' initiatives and institutional sustainability metricsModel retraining frequency ≤ monthly (for stable domains)Reduces unnecessary retraining energy cost while maintaining predictive performanceSecurity monitoring latency < 1 h for anomaly detection alertsSupports rapid response to cybersecurity and performance anomalies Representative service-level agreements (SLAs) targets and operational benchmarks for clinical deployment of in silico system (IST) Latency and responsiveness. Real-time inference is critical in high-acuity domains such as critical care or anesthesia. Practical latency targets for bedside ISTs should fall within ≤ 1–2 s for local inference and ≤ 10 s for cloud-assisted computation, with bounded maximum delay tolerances defined per use-case. For therapy-optimization or drug-dosing ISTs operating in batch or daily-update modes, latency constraints can be relaxed to minutes, but system predictability remains essential. Reliability and fault tolerance. IST systems must include automatic failover, redundancy, and clear fallback modes. For high-risk clinical functions, ≥ 99.9% uptime and deterministic fail-safes (e.g., automatic reversion to clinician judgment or default dose) are recommended. Edge inference nodes should support offline operation for ≥ 24 h to ensure continuity during network outages, with automatic state synchronization upon reconnection. Cost and resource efficiency. Model complexity should be balanced with computational feasibility. Hospitals typically operate within constrained GPU budgets and power limitations; models should support mixed-precision inference, dynamic quantization, or pruning where feasible to reduce operational cost. Use of cloud bursting for heavy training combined with local lightweight inference reduces cost and data-exposure risk. Energy efficiency and green AI. Environmental sustainability is increasingly a regulatory and institutional priority. Energy usage and compute carbon footprint should be tracked using tools such as ML CO₂ Impact or CodeCarbon, with optimization targets for model compression and data-center efficiency. Green-AI design principles (e.g., minimizing redundant retraining, using renewable-energy data centers) should be reported transparently. Integration into hospital IT infrastructure. IST deployment must comply with institutional cybersecurity, interoperability (HL7 FHIR, DICOM), and data residency policies (Fig. 5). Continuous monitoring agents should log performance metrics, failures, and computational load for traceability and service audits. Workflow for integration of in silico twin (IST) systems into hospital IT infrastructure The diagram illustrates how patient data from the electronic health record (EHR) are processed through the IST model, deployed across edge and cloud environments, and interfaced with clinical decision support systems (CDSS). Bidirectional feedback loops enable continuous learning, real-time monitoring, and model updates, ensuring traceability and adaptive personalization. Ethics and governance Dynamic consent and secondary data use In continuously learning IST frameworks, patient consent must evolve from a one-time agreement to a dynamic, ongoing process. Traditional static consent does not adequately address adaptive model updates or re-use of longitudinal data. Dynamic consent models—supported by digital interfaces—allow patients to modify or reaffirm permissions as their data are reused, aggregated, or shared for model retraining. This approach aligns with principles of transparency, autonomy, and proportionality under data protection frameworks such as the General Data Protection Regulation (GDPR). Furthermore, secondary data use for IST retraining or validation should undergo institutional review board (IRB) or ethics committee evaluation, ensuring that individual rights and public interest remain balanced. Clinician oversight and model discordance The deployment of ISTs as decision-support tools introduces ethical complexities when model-generated recommendations conflict with clinician judgment. In such cases, the clinician must retain final decision authority, supported by clear institutional protocols for documenting and explaining divergences. Explainability-by-design principles should be embedded to ensure that model outputs are interpretable, traceable, and reviewable by human experts. Ethical governance frameworks may also include AI oversight committees, multidisciplinary review boards, and audit mechanisms to monitor potential harms arising from overreliance on algorithmic outputs. These safeguards maintain accountability while supporting responsible clinical innovation. Clinical integration and workflow Translating ISTs from research prototypes into daily practice requires seamless integration with existing clinical information systems and well-defined human-in-the-loop pathways. Table 10 outlines a representative workflow across four key actors—clinician, IST platform, EHR system, and patient—showing how clinical orders, data flows, and decision loops interact in real time. Order entry and trigger. A clinician initiates a diagnostic or therapeutic request (e.g., drug dosing optimization) through the electronic health record (EHR). This action automatically triggers a secure API call to the IST platform. Order entry and trigger. A clinician initiates a diagnostic or therapeutic request (e.g., drug dosing optimization) through the electronic health record (EHR). This action automatically triggers a secure API call to the IST platform. Data ingestion and simulation. The IST platform retrieves the relevant patient data (demographics, labs, vitals, genomics, device data) via FHIR or HL7 interfaces, updates model parameters, and performs personalized simulation or prediction in near real time. Data ingestion and simulation. The IST platform retrieves the relevant patient data (demographics, labs, vitals, genomics, device data) via FHIR or HL7 interfaces, updates model parameters, and performs personalized simulation or prediction in near real time. Decision-support surface. Results are returned to the clinician within the native EHR interface—displayed as contextual decision cards summarizing predicted outcomes and recommended actions (e.g., adjusted dose, monitoring frequency, alert thresholds). Decision-support surface. Results are returned to the clinician within the native EHR interface—displayed as contextual decision cards summarizing predicted outcomes and recommended actions (e.g., adjusted dose, monitoring frequency, alert thresholds). Human-in-the-loop override. Clinicians retain full decision authority. Each recommendation includes uncertainty bounds, rationale summaries, and override options. Any clinician-initiated changes are logged for continual IST retraining and auditability. Human-in-the-loop override. Clinicians retain full decision authority. Each recommendation includes uncertainty bounds, rationale summaries, and override options. Any clinician-initiated changes are logged for continual IST retraining and auditability. Patient feedback and monitoring. When appropriate, recommendations propagate to patient-facing portals (e.g., medication schedules, wearable reminders). Data from patient sensors feed back into the IST to refine longitudinal predictions. Patient feedback and monitoring. When appropriate, recommendations propagate to patient-facing portals (e.g., medication schedules, wearable reminders). Data from patient sensors feed back into the IST to refine longitudinal predictions. Table 10Clinical integration and workflow swimlane for in silico twins (ISTs)Actor/LaneKey ActivitiesExample Interfaces/Data FlowClinician1. Initiates order entry (e.g., dosing, diagnostic request)2. Reviews IST recommendations in EHR3. Accepts or overrides recommendations4. Documents clinical decision for audit trailEHR order entry module; decision-support cards; clinician dashboardIST Platform1. Receives order trigger from EHR via API2. Ingests patient data (labs, vitals, genomics)3. Runs simulation/prediction4. Returns structured output to EHR5. Logs actions and clinician feedbackFHIR APIs; containerized model services; audit log databaseEHR System1. Transmits relevant data securely to IST platform2. Displays model outputs in clinician interface3. Records accepted/overridden decisions4. Stores results in patient recordHL7/FHIR interoperability layer; data visualization modulesPatient/Device1. Provides continuous or scheduled health data2. Receives personalized recommendations and alerts3. Sends adherence and outcome data back to IST4. Enables closed-loop learning and model updatesWearables; patient mobile app; remote monitoring dashboard Clinical integration and workflow swimlane for in silico twins (ISTs) 1. Initiates order entry (e.g., dosing, diagnostic request) 2. Reviews IST recommendations in EHR 3. Accepts or overrides recommendations 4. Documents clinical decision for audit trail 1. Receives order trigger from EHR via API 2. Ingests patient data (labs, vitals, genomics) 3. Runs simulation/prediction 4. Returns structured output to EHR 5. Logs actions and clinician feedback 1. Transmits relevant data securely to IST platform 2. Displays model outputs in clinician interface 3. Records accepted/overridden decisions 4. Stores results in patient record 1. Provides continuous or scheduled health data 2. Receives personalized recommendations and alerts 3. Sends adherence and outcome data back to IST 4. Enables closed-loop learning and model updates Case studies: oncology and cardiology Case study A — Oncology: individualized dosing and early response prediction. We illustrate a prototypical IST application in oncology where a mechanistic PK/PD backbone (tumor growth + drug effect) is personalized with patient-level imaging, pharmacokinetics, and genomic covariates to predict early response and recommend individualized dose adjustments. A pragmatic evaluation pipeline includes: Data and cohorts. Development using a retrospective cohort of ~ 1,000 treated patients (multi-site pooled EHR and imaging), internal cross-validation and temporal holdouts, and external validation on an independent cohort of 300–500 patients from geographically distinct centers. Prospective feasibility pilot: n ≈ 100–200 patients before any interventional use. Endpoints and utility targets. Primary endpoints: objective response rate (ORR) and progression-free survival (PFS) at clinically relevant timepoints (e.g., 6 months). Clinical-utility targets to justify action (dose change or escalation) are set to: absolute ORR improvement ≥ 5–10% or hazard ratio (HR) for progression ≤ 0.80 versus standard dosing, together with positive net benefit on decision-curve analysis at pre-specified risk thresholds. Safety criterion: non-inferior severe toxicity rates (grade ≥ 3 adverse events), with an acceptable risk-difference margin of ± 2–3 percentage points. Evidence and precedent. Model-informed precision dosing (MIPD) is an emerging paradigm in oncology with recent reviews summarizing methods and translational potential. Therapeutic drug monitoring (TDM) of 5-fluorouracil (5-FU) has the strongest clinical evidence among chemotherapy agents: TDM-guided 5-FU dosing has been associated with improved exposure control and reduced toxicity in several studies and reviews. Representative references include Lee et al. (Lee et al. 2016) and Gamelin et al. (Gamelin et al. 1996) on 5-FU exposure–response/toxicity relationships, and recent reviews and position pieces on MIPD (Agema et al. 2025; Maier et al. 2022). Simulation studies demonstrate feasibility of MIPD strategies for trial simulation and dosing optimization (Agema et al. 2025; Gamelin et al. 1996; Lee et al. 2016; Maier et al. 2022) (Table 10). Validation and V and V. Report discrimination (AUROC/C-index), calibration (slope/intercept), AUPRC for imbalanced outcomes, decision-curve net benefit, and NRI for risk re-classification. Mandatory external validation and subgroup analyses (histology, prior lines, performance status). Use Bayesian or ensemble UQ to provide credible intervals for recommended dose changes. Pre-specify safety stopping rules for the pilot (e.g., if grade ≥ 3 toxicity rate exceeds standard-of-care rate by a pre-set margin with p < 0.05). Key failure modes & mitigations. Cohort bias/poor representativeness — rare histologies or underrepresented ancestry groups can cause severe miscalibration. Mitigation: active external validation and re-weighted calibration; conservative thresholds for action in underrepresented subgroups. Cohort bias/poor representativeness — rare histologies or underrepresented ancestry groups can cause severe miscalibration. Mitigation: active external validation and re-weighted calibration; conservative thresholds for action in underrepresented subgroups. Missing or delayed imaging/PK data — missing modalities can prevent timely inference. Mitigation: robust imputation pipelines, modality-fallback policies, and clearly documented minimum-data requirements for actionable outputs. Missing or delayed imaging/PK data — missing modalities can prevent timely inference. Mitigation: robust imputation pipelines, modality-fallback policies, and clearly documented minimum-data requirements for actionable outputs. Out-of-distribution (OOD) treatments — novel combinations not seen during training may invalidate predictions. Mitigation: flag OOD inputs, revert to clinician judgment, and log for retraining. Out-of-distribution (OOD) treatments — novel combinations not seen during training may invalidate predictions. Mitigation: flag OOD inputs, revert to clinician judgment, and log for retraining. Uncertainty miscommunication — presenting point estimates without intervals leads to over-trust. Mitigation: mandatory uncertainty bands in decision cards and conservative recommendation language during pilots. Uncertainty miscommunication — presenting point estimates without intervals leads to over-trust. Mitigation: mandatory uncertainty bands in decision cards and conservative recommendation language during pilots. Implementation note. All numerical targets above are intended as illustrative benchmarks; authors should replace these with context-specific targets developed with domain clinicians and local regulators or with empiric effect sizes from the intended therapeutic area (Augustin et al. 2023). Oncology represents one of the most data-rich and computationally tractable domains for IST deployment. In silico tumor growth and therapy-response models have been calibrated using datasets from 200–500 patients per cancer type, integrating multi-omic, histopathologic, and imaging (CT/MRI/PET) data streams. For example, virtual twin models for breast and lung cancer incorporating radiomics and AI-driven pharmacokinetic modeling have achieved prediction accuracies of 82–90% for treatment response and progression-free survival (PFS) classification, validated in prospective cohorts (Hosny et al. 2018; Xie et al. 2021). Mechanistic–AI hybrid frameworks have also been developed for radiotherapy optimization, where dose–response ISTs reproduced observed tumor control probabilities within ± 5% error margins, outperforming conventional models in independent validation sets. Ongoing prospective trials in oncology are exploring how IST-guided adaptive therapy protocols can reduce overtreatment while preserving long-term clinical outcomes. Case study B — Cardiology: heart-failure therapy optimization and readmission prevention. A second focused IST application targets heart-failure (HF) management by combining a patient-specific physiologic model (hemodynamics, renal interplay) with remote monitoring (weight, BP, heart rate, device telemetry) and EHR data to recommend medication titration (e.g., diuretic dosing, ACEi/ARNI up-titration) and to predict 30-day readmission risk. A representative evaluation plan is: Data and cohorts. Development on ~ 5,000–10,000 longitudinal encounters (EHR + remote vitals) pooled from several hospital networks; external validation on ~ 1,000–2,000 temporally and geographically distinct encounters. Prospective pilot recommended: n ≈ 200–500 with operational endpoints. Endpoints and utility targets. Primary endpoints: 30-day all-cause readmission and time-to-first HF hospitalization. Clinically meaningful effect-size targets: relative readmission reduction 10–20% (or absolute reduction 2–5 percentage points) in high-risk subgroups, together with positive decision-curve net benefit and improved NRI. Safety monitoring focuses on hemodynamic adverse events (symptomatic hypotension) and renal dysfunction (rise in creatinine/AKI). Acceptable safety requires no significant increase in serious adverse events compared with standard care. Evidence and precedent. Hemodynamic-guided management has demonstrated clinically meaningful reductions in HF hospitalizations in randomized trials: the CHAMPION/CardioMEMS program reported substantial reductions in HF admissions (landmark results showing ~ 28–44% reductions in various analyses), and recent trials such as MONITOR-HF support pulmonary artery pressure-guided management. Non-invasive remote monitoring trials (TIM-HF2) also reported benefit (reduction in days lost due to unplanned cardiovascular admissions and mortality benefits in selected populations). These studies indicate that physiologic monitoring coupled with therapeutic management can reduce hospitalization risk and serve as realistic targets for IST-enabled interventions (Assaad et al. 2019; Cardiovascular.Abbott 2025; Koehler et al. 2024). Validation and V and V. Use time-split validation to simulate prospective deployment, subgroup calibration across age and renal function strata, and continuous post-deployment drift monitoring. Performance metrics: AUROC for binary readmission prediction, calibration plots, NRI by risk strata, and time-to-event HRs for hospitalization. Operationally critical constraints include latency targets (≤ 2 s edge inference for bedside alerts) and robust offline modes for home-monitoring interruptions. Key failure modes and mitigations. Sensor/data dropout or artifact — missing telemetry can generate false negatives/positives. Mitigation: signal-quality checks, graceful degradation, ensemble imputation, and clinician confirmation before automated actions. Sensor/data dropout or artifact — missing telemetry can generate false negatives/positives. Mitigation: signal-quality checks, graceful degradation, ensemble imputation, and clinician confirmation before automated actions. Vendor/device bias — models trained on a single device vendor’s waveform patterns may not generalize. Mitigation: multi-vendor training data, calibration layers, and external validation. Vendor/device bias — models trained on a single device vendor’s waveform patterns may not generalize. Mitigation: multi-vendor training data, calibration layers, and external validation. Alarm fatigue/spurious alerts — excessive false positives degrade clinician trust. Mitigation: conservative thresholds during pilots, aggregated alert scoring, and clinician-tunable sensitivity. Alarm fatigue/spurious alerts — excessive false positives degrade clinician trust. Mitigation: conservative thresholds during pilots, aggregated alert scoring, and clinician-tunable sensitivity. Latency/reliability issues — slow inference or outages undermine utility. Mitigation: edge inference fallback, local buffering, and strict SLA adherence. Latency/reliability issues — slow inference or outages undermine utility. Mitigation: edge inference fallback, local buffering, and strict SLA adherence. Implementation note. As above, numerical targets are illustrative; replace with local targets derived with clinicians and with evidence from published HF monitoring/titration trials where available (Bayes‐Genis et al. 2025; Givertz et al. 2017; Koehler et al. 2024). Also, quantitative validation efforts have demonstrated the growing clinical maturity of IST applications in cardiology. For example, the UVA/Padova Type 1 Diabetes Simulator, adapted for cardiovascular glucose–insulin dynamics, has been used to simulate over 300 VPs in glycemic control studies, yielding prediction accuracies above 90% for glucose variability (Kovatchev et al. 2009). In the cardiac electrophysiology domain, personalized cardiac ISTs integrating LGE-MRI–derived scar maps and CARTO voltage mapping have been validated in cohorts of 50–120 patients with ischemic cardiomyopathy, achieving area under the receiver operating characteristic (AUC) scores of 0.85–0.92 for arrhythmia localization and ablation target prediction (Prakosa et al. 2018). Additional studies using multi-physics electromechanical ISTs in heart failure have shown correlations exceeding R2 = 0.80 with measured left-ventricular ejection fraction and improved patient stratification for device therapy. Collectively, these results demonstrate the feasibility of deploying cardiac ISTs for both mechanistic understanding and patient-specific intervention planning. Model interpretability and explainability Explainable AI (XAI) is a burgeoning field focused on making complex AI models transparent, allowing stakeholders to understand how specific outputs are derived and to validate algorithmic behavior (Ray 2025). This is paramount for fostering trust and enabling the widespread adoption of AI in high-stakes clinical decision-making (Mirakhori & Niazi 2025). Transparent AI in clinical environments In clinical environments, transparent AI is crucial for building confidence among healthcare professionals. XAI provides clear explanations of the importance of various health parameters—such as patient age, gender, pre-existing conditions, and environmental factors—to different healthcare stakeholders, including drug developers, doctors, and health centers (Wang et al. 2024). Different stakeholders have distinct needs for explanation: service providers (e.g., system designers, data scientists) require logical explanations for debugging and improving model performance, while end-users (e.g., patients, non-technical clinicians, policymakers) need easy-to-understand visual explanations (Ray 2025; Wang et al. 2021). Legal auditors demand clear explanations to ensure compliance, such as avoiding racial discrimination or gender bias in loan approvals (Ray 2025). Common XAI tools include LIME (Local Interpretable Model-agnostic Explanations) and SHAP (SHapley Additive exPlanations), which convert complex feature interactions into human-understandable narratives or visualizations (Ray 2025; Wang et al. 2024). For example, in the context of diagnosing depression, XAI features (including SHAP and decision tree-based XAI) can be incorporated to help physicians understand the model’s decision-making processes and identify the factors contributing to depression presence or severity (Imans et al. 2024). It is noted that obtaining feature importance with certain dynamic ensemble selection methods can be challenging, thus the best classical ML or static ensemble model is often used to generate XAI (Imans et al. 2024). XAI also offers the most appropriate feature sets for personalized healthcare in clinical practice and enhances the transparency of AI-based illness diagnosis, drug trial result tracking, and health monitoring. Implementing XAI, however, requires careful consideration of the trade-offs between model accuracy and interpretability (Wang et al. 2024). Strengthening the V&V and UQ pipeline is necessary to bridge IST research and clinical deployment. We emphasize alignment with standards such as ASME V&V 40 for mechanistic model credibility, TRIPOD-ML and PROBAST-AI for transparent algorithm reporting, and CONSORT-AI/SPIRIT-AI for prospective clinical evaluations. Prospective pilots and sustained drift monitoring with pre-specified performance targets (e.g., NRI, decision-curve net benefit) will be critical to demonstrate both safety and clinical utility. Regulatory expectations Regulatory bodies worldwide increasingly emphasize explainability and transparency for AI in healthcare. Both the FDA (U.S.) and EMA (Europe) prioritize ethical considerations, transparency, and patient data privacy (e.g., under GDPR) in their regulatory approaches to AI in healthcare (Kumar et al. 2025; Mirakhori & Niazi 2025). The General Data Protection Regulation (GDPR) notably highlights a "right to explanation," driving the demand for greater interpretability in AI systems (Ghorbian et al. 2025; Wang et al. 2021). The FDA's guidance on AI in medical device development similarly stresses the need for transparency and validation (Mirakhori & Niazi 2025). Establishing standardized protocols for model verification, validation, and uncertainty quantification (VV&UQ), aligned with guidelines from bodies like the FDA or EMA, is a projected milestone for achieving regulatory-grade model certification (Table 11) (Xu & Wang 2025). This ensures that AI-driven software tools can be classified as Software as a Medical Device (SaMD), requiring developers to meet rigorous validation, transparency, and traceability standards (Xu & Wang 2025). Table 11Challenges and current strategies for validation and regulatory alignment in in silico healthcare modeling. Iterative mitigation approaches and evidence types noted. Representative references: (Abdollahi et al. 2024); (Viceconti et al. 2021); (Ferle et al. 2025). Legend: data flow = challenge identification → mitigation approach → governance/monitoring loop. Update cadence: iterative and continuous as standards and evidence evolveDomain/SystemCore Focus/AspectModel Type/RoleValidation Dataset/TrialEvidence TypeKey Outcomes/Success IndicatorsRepresentative ReferencesValidation & GeneralizabilityLack of external validation; inconsistent evaluation protocolsMethodological frameworks (VV&UQ, silent-mode trials)Varied; multi-center external cohortsRetrospective & MethodologicalAdoption of external validation and standardized protocols(Ferle et al. 2025; Viceconti et al. 2021)Data QualityHeterogeneous sources, demographic gapsData harmonization, federated learningMulti-institutional datasetsRetrospectiveImproved robustness and privacy-preserving multi-site training(Abdollahi et al. 2024; Ghorbian et al. 2025)Explainable AILack of XAI metrics and stakeholder-specific explanationsSHAP, LIME, model cards, user-centered interfacesEvaluation studiesMethodological/RetrospectiveBetter clinician trust and auditability(Ray 2025; Wang et al. 2024) Challenges and current strategies for validation and regulatory alignment in in silico healthcare modeling. Iterative mitigation approaches and evidence types noted. Representative references: (Abdollahi et al. 2024); (Viceconti et al. 2021); (Ferle et al. 2025). Legend: data flow = challenge identification → mitigation approach → governance/monitoring loop. Update cadence: iterative and continuous as standards and evidence evolve Challenges and limitations The promising potential of ISTs in precision medicine is accompanied by several significant challenges that must be thoroughly addressed before their widespread adoption in daily clinical settings (Ștefănigă et al. 2024). These challenges span technical, ethical, computational, and clinical domains, necessitating multidisciplinary collaboration and robust frameworks for their effective integration into healthcare (Table 12) (Ajax & Gimah 2025; Javed et al. 2024; Owolabi & Owolabi 2025; Ștefănigă et al. 2024; Tiwari et al. 2025; Vengathattil 2025). Table 12Technical, ethical, and clinical challenges in deploying in silico twins with proposed mitigation strategies. Evidence types and cadence included. Representative references: (Ajax & Gimah 2025); (Badidi 2023); (Javed et al. 2024). Legend: data flow = development → deployment → post-market monitoring; Update cadence: development continuous, post-market monitoring per regulatory cyclesDomain/SystemCore Focus/AspectModel Type/RoleValidation Dataset/TrialEvidence TypeKey Outcomes/Success IndicatorsRepresentative ReferencesTechnical DeploymentInteroperability, computational cost, model instabilityStandards (HL7 FHIR), edge/cloud hybridization, surrogate modelsPilot deploymentsMixedScalable deployment pathways and standardized interfaces(Ajax & Gimah 2025; Badidi 2023)Ethical & PrivacyData privacy, ownership, liabilityFederated learning, encryption, governance boardsRegulatory guidance casesPolicy/RetrospectiveCompliance with GDPR/HIPAA and patient consent workflows(Kumar et al. 2025; Mirakhori & Niazi 2025)Clinical IntegrationWorkflow fit, clinician trust, costHuman-in-the-loop, XAI, training programsClinical pilotsProspective/PilotImproved clinician acceptance and task efficacy(Javed et al. 2024; Tiwari et al. 2025) Technical, ethical, and clinical challenges in deploying in silico twins with proposed mitigation strategies. Evidence types and cadence included. Representative references: (Ajax & Gimah 2025); (Badidi 2023); (Javed et al. 2024). Legend: data flow = development → deployment → post-market monitoring; Update cadence: development continuous, post-market monitoring per regulatory cycles Data privacy and ethical challenges The use of ISTs and AI in healthcare faces substantial data privacy and ethical challenges due to the use of sensitive multimodal patient data (Alsaedi et al. 2025; Badidi 2023; Garg 2025; Haferlach et al. 2025; Javed et al. 2024; Kant And Deepika 2025; Leivaditis et al. 2025; Ocana et al. 2025; Owolabi & Owolabi 2025; Panchpuri et al. 2025; Ștefănigă et al. 2024; Tiwari et al. 2025; Vengathattil 2025). (Garg 2025) highlights the risk of re-identification even in synthetic data, and bias arising from underrepresented populations. Pishvaian et al. (Pishvaian et al. 2019) addressed this by developing a secure virtual tumor board, reducing turnaround time from 14 to 4 days for treatment reports while maintaining compliance (Fig. 6A). Despite these advances, broader deployment still requires robust anonymization, clear consent frameworks, and methods to audit fairness and ethical compliance in model outputs. Moreover, managing vast amounts of sensitive personal data—including genetic and lifestyle information—raises concerns about its security, storage, and potential misuse by third parties, including ransomware risks (Alsaedi et al. 2025; Badidi 2023; Garg 2025; Javed et al. 2024; Kalusivalingam et al. 2024; Panchpuri et al. 2025; Ștefănigă et al. 2024; Tiwari et al. 2025; Vengathattil 2025). Adherence to stringent data protection regulations like GDPR and HIPAA is crucial but often complex (Ajax & Gimah 2025; Haferlach et al. 2025; Panchpuri et al. 2025; Ștefănigă et al. 2024). A significant ethical concern is algorithmic bias, where models trained on imbalanced datasets can perpetuate inequalities and lead to unfair or inaccurate outcomes, particularly for underrepresented patient groups (Ajax & Gimah 2025; Alsaedi et al. 2025; Javed et al. 2024; Leivaditis et al. 2025; Panchpuri et al. 2025; Ștefănigă et al. 2024; Tiwari et al. 2025; Vengathattil 2025). Ensuring accountability for AI-driven decisions and obtaining fully informed patient consent are also paramount (Ajax & Gimah 2025; Badidi 2023; Garg 2025; Haferlach et al. 2025; Javed et al. 2024; Ștefănigă et al. 2024; Vengathattil 2025). Early involvement of ethicists is vital to foster trust and ensure transparency (Ștefănigă et al. 2024; Vengathattil 2025). A Workflow of the virtual molecular tumor board (VMTB) system developed by Perthera for personalized oncology. Data flow begins with patient enrollment and record aggregation, proceeds through molecular analysis and AI-assisted therapy ranking, and concludes with clinician consensus and feedback integration. Update cadence: case-level updates with continuous registry augmentation (prospective). (B) Relative prevalence of generative AI model use in medical applications, highlighting GAN and VAE architectures. Evidence type: Prospective clinical registry (A) and literature-derived data synthesis (B).* Panel A reproduced with permission from Pishvaian et al. (Pishvaian et al. 2019), Copyright 2019 Oxford University Press; Panel B reproduced with permission from Garg (Garg 2025), Copyright 2025 Garg. Abbreviations: GAN, generative adversarial network; VAE, variational autoencoder; EHR, electronic health record Interoperability and data silos A significant hurdle for ISTs and AI in healthcare is the heterogeneity and fragmentation of patient data, often trapped in "data silos" (Ajax & Gimah 2025; Badidi 2023; Garg 2025; Javed et al. 2024; Ocana et al. 2025; Owolabi & Owolabi 2025; Ștefănigă et al. 2024; Tiwari et al. 2025; Ullah et al. 2024). In this regard, Garg (Garg 2025) notes that most generative AI systems process only one data type. Pishvaian et al. (Pishvaian et al. 2019) demonstrated partial success in integrating molecular, treatment, and clinical records, yet only 1,725 patients were served over 4 years. Cerrone et al. (Cerrone et al. 2025) relied solely on neuroimaging for model input. EHRs, imaging, and wearable data frequently use differing formats and protocols, impeding data integration, harmonization, and standardization (Ajax & Gimah 2025; Badidi 2023; Garg 2025; Haferlach et al. 2025; Javed et al. 2024; Owolabi & Owolabi 2025; Ștefănigă et al. 2024; Tiwari et al. 2025; Vengathattil 2025). This lack of interoperability obstructs timely, comprehensive access to critical patient information, impacting continuity of care and effective AI application (Badidi 2023; Owolabi & Owolabi 2025). AI algorithms trained on one system's data may perform poorly in another without standardization (Tiwari et al. 2025). Moreover, sharing sensitive health data across diverse systems without proper frameworks poses privacy and security risks (Badidi 2023; Javed et al. 2024). Solutions involve developing standardized formats (e.g., HL7 FHIR, OMOP), middleware, and fostering collaborative data sharing among institutions (Badidi 2023; Javed et al. 2024; Owolabi & Owolabi 2025; Ștefănigă et al. 2024; Vengathattil 2025) because without standardized frameworks like FHIR or OMOP, data silos will continue to restrict model scalability and real-time adaptability. Scalability and computational demand Developing and maintaining ISTs demands significant computational, human, and financial resources (Ștefănigă et al. 2024; Vengathattil 2025). Training and deploying AI models, especially generative AI, require high-performance computing (HPC) infrastructure, often leveraging cloud or edge computing (Ajax & Gimah 2025; Alsaedi et al. 2025; Badidi 2023; Garg 2025; Javed et al. 2024; Ocana et al. 2025; Panchpuri et al. 2025; Ștefănigă et al. 2024; Vengathattil 2025). Enormous genomic datasets, crucial for precision medicine, necessitate advanced computational power for analysis (Ajax & Gimah 2025). The constant need for real-time data integration and updates for ISTs creates substantial technical barriers and latency concerns (Alsaedi et al. 2025; Garg 2025; Vengathattil 2025). The high costs of hardware, training costs, and infrastructure, coupled with the need for specialized technical expertise, present a major hurdle for widespread implementation, particularly for smaller healthcare institutions (Ajax & Gimah 2025; Ștefănigă et al. 2024; Vengathattil 2025). Optimizing model architectures and utilizing cloud-based AI acceleration are crucial mitigation strategies (Alsaedi et al. 2025; Badidi 2023; Garg 2025; Javed et al. 2024; Vengathattil 2025). In this regard, Cerrone et al. (Cerrone et al. 2025) achieved a 10 × speedup using reduced-order modeling for glioblastoma simulation, enabling real-time prediction. However, model training remained computationally heavy. Kalusivalingam et al. (Kalusivalingam et al. 2024) demonstrated that deep learning models achieved a 15% improvement over mutual information (MI) in diagnostic accuracy for image registration, with enhancements in sensitivity from 0.85 to 0.92 and specificity from 0.88 to 0.94. The effectiveness of registration was primarily assessed using target registration error (TRE). The MI method yielded an average TRE of 2.8 mm on the test samples, while the deep learning models produced a significantly reduced TRE of 1.2 mm, indicating a 57% enhancement compared to the MI approach. This difference is statistically significant, with a p-value of less than 0.05, as shown by paired t-tests across different image modalities. Garg (Garg 2025) examined how generative AI techniques, including GANs, VAEs, and transformer-based architectures like GPT, are applied in personalized treatment planning, drug discovery, and synthetic data generation (Fig. 6B). The study highlighted that methods such as transformer architectures, GANs, and VAEs are particularly potent in drug discovery and molecular design. These models are adept at analyzing extensive chemical and biological datasets to identify innovative compounds and accurately predict their interactions. This capability significantly shortens the drug development timeline, reducing it from over a decade to just a few months or years. Clinical trust and interpretability A major impediment to adopting ISTs and AI in clinical settings is the lack of transparency and interpretability of many AI models, often termed "black boxes" (Ajax & Gimah 2025; Alsaedi et al. 2025; Badidi 2023; Garg 2025; Haferlach et al. 2025; Javed et al. 2024; Leivaditis et al. 2025; Panchpuri et al. 2025; Ștefănigă et al. 2024; Tiwari et al. 2025; Vengathattil 2025). Garg (Garg 2025) criticizes black-box AI for lacking clinician trust. Healthcare professionals must understand and trust how AI systems reach decisions to integrate them into workflows and maintain accountability for patient outcomes (Ajax & Gimah 2025; Alsaedi et al. 2025; Garg 2025; Haferlach et al. 2025; Leivaditis et al. 2025; Panchpuri et al. 2025; Tiwari et al. 2025). This interpretability deficit hinders clinician acceptance and patient trust (Ajax & Gimah 2025; Alsaedi et al. 2025; Garg 2025; Tiwari et al. 2025). Kalusivalingam et al. (Kalusivalingam et al. 2024) noted deep learning improved accuracy but was less interpretable than MI. Cerrone et al. (Cerrone et al. 2025) addressed this through sensitivity analysis, helping explain tumor model behavior. Pishvaian et al. (Pishvaian et al. 2019) relied on expert-reviewed treatment plans via a virtual board to ensure trust. Furthermore, rigorous clinical validation is essential to establish efficacy, safety, and reliability in diverse patient populations, a crucial step often lacking in current research (Ajax & Gimah 2025; Alsaedi et al. 2025; Badidi 2023; Leivaditis et al. 2025; Ștefănigă et al. 2024; Tiwari et al. 2025). Developing Explainable AI (XAI) techniques and conducting extensive validation studies are crucial to building confidence and accelerating adoption (Ajax & Gimah 2025; Garg 2025; Haferlach et al. 2025; Javed et al. 2024; Tiwari et al. 2025). Ultimately, blending XAI models with clinician oversight may best support safe and confident use in clinical settings. Ethical, regulatory, and legal perspectives Medical digital twin (MDT) technology, integral to precision medicine, promises advanced diagnosis, treatment simulation, and drug discovery by considering individual variability in genes, environment, and lifestyle (Shao & Ou 2025; Tian et al. 2025). However, its application introduces significant ethical, regulatory, and legal challenges (Apu et al. 2022; Chassang et al. 2025; Ghosh & Bhattacharjee 2024; Strigari et al. 2025). These concerns include data ownership, patient consent, liability in simulated decisions, and compliance with existing regulations. The reliance on extensive physical data raises questions about privacy, potential algorithmic bias, and fairness in treatment recommendations (Apu et al. 2022; Shao & Ou 2025; Strigari et al. 2025). The technology's ability to comprehensively monitor and influence patient behavior can lead to issues of data centralization and a new form of biopolitical control, described as a "Leviathan" (Shao & Ou 2025). Addressing these requires developing specific legal frameworks, ensuring regulatory harmonization across jurisdictions, and fostering interdisciplinary collaboration to balance innovation with patient safety and privacy (Apu et al. 2022; Chassang et al. 2025; Prifti et al. 2024; Strigari et al. 2025). Informed consent and digital ethics Informed consent is a critical ethical consideration for ISTs in precision medicine, particularly due to the continuous collection and sharing of sensitive patient data, including imaging and longitudinal clinical measures (Chassang et al. 2025; Strigari et al. 2025). Patients cede their bodily data rights under data-sharing contracts to enable the construction and operation of their ISTs, raising concerns about data ownership and privacy (Shao & Ou 2025). Shao and Ou (Shao & Ou 2025) emphasize the bio-political risk of patient digital surveillance, highlighting that DTs could lead to subtle coercion or behavioral control under the guise of care optimization. This aligns with the concerns raised by Chassang et al. (Chassang et al. 2025), who advocate for operational ethics frameworks to protect individual autonomy in AI-driven public health systems. Current broad consent models are often insufficient for the "secondary use" of physiological data across various fields like treatment, academic research, and enterprises (Shao & Ou 2025). A dynamic consent approach is proposed, requiring re-consent when the data application link changes, empowering patients to monitor and terminate data usage (Shao & Ou 2025). Balancing data utilization for intelligent healthcare with robust privacy protection is essential, necessitating de-identification and the use of privacy-preserving technologies such as FL or homomorphic encryption (Apu et al. 2022; Chassang et al. 2025; Shao & Ou 2025). The literature converges on the necessity of dynamic informed consent protocols—iterative and adaptive to the evolving scope of data use—alongside a clear delineation of patient rights and revocation mechanisms. Quantitatively, over 70% of surveyed students and healthcare experts in Choi et al. (2024) agreed on the need for ethical digital health training, reinforcing the practical urgency for institutional safeguards around consent and digital rights. Despite anonymization, re-identification risks persist, highlighting the need for a hierarchical system of dynamic protection obligations for personal information and continuous security assessments (Shao & Ou 2025; Strigari et al. 2025). Regulatory pathways for clinical twin approval (FDA, EMA frameworks) The regulation of clinical ISTs falls under emerging frameworks for AI systems and medical devices (Chassang et al. 2025). In the European Union (EU), the Artificial Intelligence Act (AIA) is a horizontal regulation that interacts with the Medical Device Regulations (MDR) and the General Data Protection Regulation (GDPR) to govern AI applications, including those in healthcare (Chassang et al. 2025). These regulations aim to ensure trustworthiness in AI by requiring technical robustness, reliability, and security. Mandatory assessments, such as the Fundamental Right Impact Assessment (FRIA) for high-risk AI systems, are crucial before deployment (Chassang et al. 2025). Regulatory pathways for clinical twin approval will necessitate rigorous testing and validation through clinical trials to ensure accuracy, reliability, and safety (Chassang et al. 2025; Strigari et al. 2025). Strigari et al. (Strigari et al. 2025) argue that robust clinical deployment of theranostic ISTs requires harmonized protocols across institutions and borders—highlighting that current legislation limits data interoperability, particularly due to divergent GDPR and HIPAA interpretations. Chassang et al. (Chassang et al. 2025) analyze frameworks from UNESCO, WHO, and the EU AI Act, noting that although ethical principles are widely recognized (e.g., fairness, transparency, non-maleficence), enforcement mechanisms are often vague or underdeveloped. Quantitative gaps remain: only 18% of surveyed stakeholders reported clarity on liability or safety regulation routes for AI-driven tools. This highlights the need for updated regulatory guidelines explicitly addressing IST verification, model explainability, and cross-jurisdictional data governance. The development of technical standards and certifications for AI systems, especially those not classified as medical devices, is also vital (Chassang et al. 2025). Globally, regulatory harmonization is needed to address fragmentation and ensure consistent standards for cross-border deployment and data transfer (Apu et al. 2022). Liability and safety concerns in simulated decisions ISTs offer capabilities for scenario analysis, predictive analytics, and clinical decision support, aiming to achieve "zero-error patient-care" by identifying and mitigating risks proactively (Ghosh & Bhattacharjee 2024; Shao & Ou 2025). However, integrating ISTs raises significant liability and safety concerns, particularly in decisions influenced by simulations (Ghosh & Bhattacharjee 2024; Strigari et al. 2025). Potential risks include technical inaccuracies in computational models, data privacy breaches, and clinical errors from incorrect treatment recommendations (Strigari et al. 2025). The "black-box" nature of AI algorithms can lead to opacity, bias, and a lack of interpretability, undermining user trust and complicating accountability (Apu et al. 2022; Shao & Ou 2025). Medical ISTs enable comprehensive monitoring, creating a "digital panopticon" where patient data is continuously analyzed, leading to concerns about concealed control and the alienation of individual freedom (Shao & Ou 2025). Establishing clear liability frameworks and robust risk management strategies, including continuous monitoring and validation of TDT outputs, is crucial to prioritize patient safety and treatment efficacy (Strigari et al. 2025). Liability in AI-driven healthcare simulations is one of the most pressing unresolved issues. Strigari et al. (Strigari et al. 2025) point out that the ownership of IST models—especially those involving multi-institutional inputs—complicates accountability in adverse outcomes. The absence of clearly defined risk attribution mechanisms, particularly when models fail due to incomplete or outdated input data, poses challenges for both clinicians and developers. Legal analyses by Chassang et al. (Chassang et al. 2025) stress the need for regulatory reform that includes post-deployment monitoring and real-time auditing to trace accountability. Furthermore, Shao and Ou (Shao & Ou 2025) raise philosophical concerns that ISTs may shift decision-making power away from clinicians, diffusing liability. Quantitatively, fewer than 25% of stakeholders surveyed in the reviewed studies could identify the responsible party in a hypothetical scenario involving IST model error, indicating critical gaps in current legal and institutional clarity. Future directions and research priorities The advancement of ISTs in precision medicine necessitates addressing several key challenges and exploring promising research avenues. These include integrating real-time data from advanced sensors, leveraging decentralized learning paradigms, developing comprehensive personalized health maintenance models, and expanding the scope of twin models to encompass intricate biological systems like metabolism, the immune system, and the microbiome. Overcoming these hurdles will be crucial for realizing the full transformative potential of ISTs in healthcare (Haykal 2025; Sigawi & Ilan 2023). Moreover, the translational potential of ISTs remains tempered by the scarcity of open, harmonized datasets for model calibration and validation; the limited number of prospective, specialty-specific pilot studies; and the absence of widely accepted V&V and governance templates. Furthermore, long-term monitoring of model drift and bias remains underdeveloped, as do standardized procedures for post-deployment auditing. Integration with real-time biosensors and digital phenotyping The future of ISTs in precision medicine heavily relies on the integration of real-time biosensors and digital phenotyping technologies (Haykal 2025; Petrušić 2025). The integration of ISTs with real-time biosensors and digital phenotyping tools can significantly enhanced their adaptability and clinical relevance. Wearable devices (e.g., ECG monitors, glucose sensors, and activity trackers) and environmental sensors can continuously collect granular, multimodal data on a patient's physiological state, lifestyle factors, and environmental influences, which is critical for creating dynamic and accurate digital replicas (Ali et al. 2024; Ghatti et al.2023); Roopa & Venugopal 2025; Xames & Topcu 2025; Zhang et al. 2024). Wearable technologies enable continuous data flow into the IST model, facilitating dynamic updates that reflect the patient’s evolving physiological state. This continuous influx of data allows for real-time monitoring, enabling early detection of health issues and timely interventions (Ghatti 2023; Xames & Topcu 2025). Advanced AI-driven analytics can then process this rich, longitudinal data, enhancing predictive accuracy for disease progression, optimizing treatment plans, and even identifying pre-illness conditions (Ghatti 2023; Khan et al. 2022b, c, a; Petrušić 2025; Roopa & Venugopal 2025; Sigawi & Ilan 2023). Addressing the challenges of continuous data collection and integration will be vital for real-time adaptability and scalability (Roopa & Venugopal 2025). Liu et al. (Liu et al. 2025) utilized wearable and genetic data from the Adolescent Brain Cognitive Development (ABCD) study to investigate the role of digital phenotypes in psychiatric research. Drawing on more than 250 wearable-derived features, the authors demonstrated that an interpretable AI framework can classify adolescents with psychiatric disorders with improved accuracy compared to existing methods. To link these digital phenotypes to genetic underpinnings, they integrated them into both univariate and multivariate genome-wide association studies (GWAS), identifying 16 significant loci and 37 psychiatric-associated genes, including ELFN1 and ADORA3. The findings highlight that continuous, wearable-derived features provide greater statistical power than conventional case–control GWAS, underscoring the potential of wearable technologies to reveal novel connections between behavior and genetics. Federated learning for decentralized twin training FL is a crucial direction for enabling decentralized training of ISTs, addressing paramount concerns regarding data privacy, institutional data sovereignty, and security in healthcare (Fonseca et al. 2024; Sigawi & Ilan 2023; Xames & Topcu 2025; Zhang et al. 2024; Zhang et al. 2025). This approach allows ML models to be trained locally at various medical institutions, sharing only aggregated model updates—rather than sensitive patient data—with a central server (Xames & Topcu 2025; Zhang et al. 2024). Moreover, unlike traditional centralized models, FL enables collaborative learning across multiple clinical sites without the need to transfer sensitive patient data, thereby aligning with global data protection regulations such as the GDPR. This horizontal FL facilitates collaborative learning across diverse datasets while ensuring patient confidentiality and data integrity (Xames & Topcu 2025; Zhang et al. 2024). Implementing FL, potentially in conjunction with blockchain technology, can enhance the security, traceability, and tamper-proof nature of healthcare data transactions within IST frameworks (Ghatti 2023; Petrušić 2025; Roopa & Venugopal 2025; Xames & Topcu 2025; Zhang et al. 2025). This decentralized paradigm is essential for leveraging large, multi-institutional datasets without compromising privacy (Zhang et al. 2024). Personalized health maintenance twins Personalized health maintenance twins represent a significant future direction, focusing on proactive, individualized care and optimal health outcomes (Alavi et al. 2024a, b, c, d; Roopa & Venugopal 2025). Health maintenance twins represent a novel application of IST technology focused on chronic disease management and preventive care. These twins, driven by predictive modeling and multi-criteria decision-making, can simulate the effects of various interventions, from therapeutic regimens to lifestyle modifications (Roopa & Venugopal 2025; Sigawi & Ilan 2023; Zhang et al. 2024; H. Zhu et al. 2025a, b). These models are continuously updated with multimodal data streams—such as symptom reports, medication adherence, lifestyle metrics, and environmental exposures—to provide early warnings and support self-management strategies. They enable the testing of personalized treatment plans before application to the actual patient, leading to optimized and tailored care (Alavi et al. 2024a, b, c, d; Ghatti 2023; Roopa & Venugopal 2025; Zhang et al. 2024). A key aspect involves developing dynamic IST models that continuously evolve with changes in a patient's health status, accounting for aging, disease progression, and treatment responses (Roopa & Venugopal 2025). This requires incorporating concepts such as personalized variability signatures and second-generation AI systems that utilize inherent biological noise to improve accuracy and long-term efficacy (Sigawi & Ilan 2023). Such twins will offer comprehensive health guidance, enabling a shift towards preventative and adaptive healthcare (Roopa & Venugopal 2025; Xames & Topcu 2025; Zhang et al. 2024). Next-gen twin models: metabolic, immune system, microbiome modeling Developing next-generation IST models will involve delving deeper into complex biological systems, including metabolic pathways, the immune system, and the microbiome (Fonseca et al. 2024; Manchel et al. 2024). These advanced models require explicit incorporation of and driving by single-cell omics datasets and spatial resolution data, bridging the current gap between data-driven and mechanistic models (Manchel et al. 2024). These next-generation twins aim to capture the interplay of biological subsystems, enabling systems-level simulation of disease progression and therapeutic response. This will allow for granular understanding of cellular states, interactions, and trajectories within tissues and organs (Manchel et al. 2024). Furthermore, the creation of multi-organ and even whole-body ISTs will enable comprehensive simulations of system-level pathophysiology and drug interactions (Fonseca et al. 2024; Manchel et al. 2024; Roopa & Venugopal 2025; Zhang et al. 2024). Overcoming computational challenges related to high-dimensionality and integrating diverse data modalities will be key to unlocking their potential for drug discovery, treatment optimization, and enhancing the mathematical understanding of biological mechanisms (Fonseca et al. 2024; Ghatti 2023; Manchel et al. 2024; Sigawi & Ilan 2023). Moreover, to advance the field from conceptual frameworks to validated clinical tools, a focused roadmap for the next 12–24 months can be proposed as follow and box 1: Benchmarking and open datasets: Establish shared, anonymized benchmark datasets across cardiology, oncology, and metabolic disease domains to enable reproducibility and performance comparison. Benchmarking and open datasets: Establish shared, anonymized benchmark datasets across cardiology, oncology, and metabolic disease domains to enable reproducibility and performance comparison. Prospective pilots: Conduct real-world, prospective validation studies in at least two clinical specialties, with harmonized outcome measures and pre-registered protocols. Prospective pilots: Conduct real-world, prospective validation studies in at least two clinical specialties, with harmonized outcome measures and pre-registered protocols. Standardized V&V frameworks: Develop consensus-driven V&V checklists and reference workflows aligned with FDA, EMA, and SFDA regulatory expectations. Standardized V&V frameworks: Develop consensus-driven V&V checklists and reference workflows aligned with FDA, EMA, and SFDA regulatory expectations. Governance templates: Publish and pilot open governance templates covering data provenance, auditability, consent management, and post-deployment monitoring. Governance templates: Publish and pilot open governance templates covering data provenance, auditability, consent management, and post-deployment monitoring. Collectively, these steps will enable reproducible, ethically governed, and clinically credible IST development, while supporting interoperability and trust within digital health ecosystems. Box 1. Minimal reporting items recommended for transparent, reproducible, and clinically oriented in silico twin (IST) studies. ItemDescription/Guidance1. Problem definitionClearly specify the clinical question, disease domain, and intended use2. Data sources & freshnessDescribe all input datasets, their origin, age, update frequency, and curation processes3. Mechanistic–ML couplingExplain how mechanistic and AI/ML components are integrated, synchronized, and validated4. Calibration & external validationReport dataset size, diversity, and independence for calibration and external validation5. Uncertainty Quantification (UQ)Provide quantitative uncertainty or sensitivity analyses to express model confidence6. Monitoring & drift detectionOutline procedures for post-deployment monitoring, retraining, and drift detection7. Governance & documentationDetail ethical oversight, data access control, audit logs, and compliance with GDPR or similar frameworks8. Fairness & equity metricsInclude bias detection, subgroup analysis, and fairness evaluation across demographic strata9. Clinically actionable outputsDescribe how outputs support clinical decision-making, interpretability, and user communication10. Interpretability & ExplainabilityDescribe methods (e.g., SHAP, LIME) used to interpret model outputs and confirm alignment with physiological mechanisms11. Bias & Fairness MonitoringReport fairness metrics, demographic balance, and drift detection procedures across patient cohorts In addition, a Table of open resources (datasets, simulators, ontologies, benchmarks) to help reproducibility can be as follow (Table 13): Table 13Open-access datasets, simulators, ontologies, and benchmark frameworks supporting reproducible and transparent IST researchCategoryName/ResourceType/DomainDescription/ApplicationAccess/LinkDatasetsMIMIC-III/MIMIC-IVCritical Care/EHRLarge-scale, de-identified ICU data with vitals, lab results, and free-text notes for model validationhttps://mimic.mit.eduDatasetsUK BiobankPopulation Health500,000-participant cohort with imaging, genomics, and EHR linkage for AI and mechanistic modelinghttps://www.ukbiobank.ac.ukDatasetsTCGA/TCIAOncologyThe Cancer Genome Atlas & Imaging Archive; multi-omic and imaging data for tumor modelinghttps://portal.gdc.cancer.govSimulatorsUVA/Padova T1D SimulatorMetabolic/DiabetesFDA-approved in silico model for Type 1 Diabetes control algorithm testinghttps://www.est.di.unimi.itSimulatorsOpenCORMultiscale PhysiologyOpen-source modeling environment for systems biology and cardiac electrophysiologyhttps://www.opencor.wsSimulatorsPhysiome Project/VPHSystems PhysiologyEU-supported initiative providing multi-scale human physiology models and datasetshttps://physiomeproject.orgOntologies & StandardsSNOMED CTClinical TerminologyGlobal standard for encoding clinical terms in interoperable EHR-linked ISTshttps://www.snomed.orgOntologies & StandardsOBO Foundry OntologiesBiomedical OntologiesCommunity-driven biological and clinical ontology repository (Gene, Disease, Cell, etc.)https://obofoundry.orgOntologies & StandardsPhysioNetPhysiological SignalsRepository for ECG, EEG, and physiological datasets used in cardiac and neurological ISTshttps://physionet.orgBenchmarks & FrameworksFDA In silico Trials Benchmark (ISCT)Regulatory ValidationOpen framework for model qualification and comparison under FDA guidancehttps://isct.fda.govBenchmarks & FrameworksSBML/CellML ModelsSystems BiologyCommunity-curated model formats enabling reproducible exchange of biological modelshttps://sbml.orgBenchmarks & FrameworksBioModels DatabaseMechanistic ModelsRepository of published computational models for pharmacokinetics and QSPhttps://www.ebi.ac.uk/biomodels Open-access datasets, simulators, ontologies, and benchmark frameworks supporting reproducible and transparent IST research Conclusion ISTs represent a paradigm shift in precision medicine, offering a comprehensive and dynamic approach to modeling human physiology, simulating disease progression, and personalizing therapy. Unlike static models or traditional digital health tools, ISTs integrate real-time multimodal data with advanced AI, physiological modeling, and simulation engines to mirror a patient’s health state continuously. This adaptive capability enables clinicians and researchers to anticipate disease trajectories, test interventions virtually, and deliver data-driven decisions tailored to individual profiles. The convergence of AI, cloud-edge computing, and high-resolution biosensors is accelerating the development and deployment of ISTs across diverse clinical domains. Applications now span the patient care continuum—from preventive risk assessment and early diagnosis to personalized treatment, monitoring, and long-term disease management. As these systems become increasingly embedded within clinical infrastructure, they have the potential to transform not only diagnostics and therapeutics but also clinical trial design, regulatory science, and health system optimization. Despite significant advancements, challenges remain in standardizing data integration, validating predictive performance, and ensuring ethical and regulatory alignment. However, with ongoing innovation in computational modeling, federated data ecosystems, and explainable AI, the foundation for scalable, equitable, and trustworthy IST systems is rapidly maturing. Ultimately, ISTs hold the promise to redefine medical practice, not merely by digitizing the patient, but by creating an intelligent, evolving companion capable of guiding precision care in real time. Their adoption marks a critical step toward the future of individualized, proactive, and simulation-driven healthcare. Acknowledgements The author would like to thank the Deanship of Scientific Research at Shaqra University for supporting this work. Abbreviations Adolescent Brain Cognitive Development Angiotensin-Converting Enzyme inhibitor/Angiotensin Receptor-Neprilysin Inhibitor Acute Kidney Injury Artificial Intelligence Artificial Intelligence Act Artificial Intelligence–enabled Clinical Simulation and Reasoning platform Application Programming Interfaces Augmented Reality Autism spectrum disorder American Society of Mechanical Engineers' standard for Verification and Validation Area under the precision–Recall Curve Area Under the Receiver Operating Characteristic Curve Biology-Informed Neural Networks Blood Pressure Concordance Index Chimeric Antigen Receptor (CAR) T cell Three-Dimensional Electroanatomical Mapping System (Biosense Webster, USA) Clinical Decision Support System Computational Fluid Dynamics CardioMEMS Heart Sensor Allows Monitoring of Pressure to Improve Outcomes in NYHA Class III Heart Failure Patients Continuous Integration/Continuous Deployment Convolutional Neural Networks CONsolidated Standards Of Reporting Trials—Artificial Intelligence Coronavirus Disease 2019 Cardiac Resynchronization Therapy Computed Tomography Digital Imaging and Communications in Medicine Digital twin Electrocardiogram Electronic Health Records Electrical Impedance Tomography European Medicines Agency Electromyography Extract, Transform, Load European Union Findable, Accessible, Interoperable, and Reusable U.S. Food and Drug Administration Fast Healthcare Interoperability Resources Federated Learning Fundamental Right Impact Assessment Generative Adversarial Networks General Data Protection Regulation (European Union data protection and privacy law) Generative Artificial Intelligence General Electric HealthCare Good Machine Learning Practice Generative Pre-trained Transformer Graphics Processing Unit Graphical User Interface Genome-Wide Association Studies Health Economics Heart-Failure Health Insurance Portability and Accountability Act Human Immunodeficiency Virus Health Level Seven, High-Performance Computing Hazard Ratio Hardware Security Modules Health Technology Assessment Identity and Access Management Intensive Care Unit Ischemic Heart Disease Internet of Medical Things Institutional Review Board In silico Twin Information Technology Late Gadolinium Enhancement Magnetic Resonance Imaging Local Interpretable Model-agnostic Explanations Large Language Model Meta AI, version 3 (open-weight large language model developed by Meta AI) Large Language Models Low-Rank Adapters Last Observation Carried Forward Mean Absolute Error Medical Device Regulations Medical Digital Twin Mechano-Electrical-Fluid Interaction Micro-Electro-Mechanical Systems Model-Informed Drug Development Medical Information Mart for Intensive Care III (de-identified critical-care Database Developed by the MIT Laboratory for Computational Physiology) Model-Informed Precision Dosing Maximum Voluntary Isometric Contraction Machine Learning Multicentre Randomised Controlled Trial of Remote Monitoring for Heart Failure Using a Implanted Pulmonary Artery Pressure Sensor in the Netherlands Mater Private Hospital Magnetic Resonance Imaging Computational Modeling & Simulation Mean Squared Error Mutual Information National Aeronautics and Space Administration Net Reclassification Improvement Observational Medical Outcomes Partnership Out-of-Distribution Objective Response Rate Physiologically Based Pharmacokinetic Physiological Cross-Sectional Area Predetermined Change Control Plan Programmed Death Ligand 1 Progression-Free Survival Personally Identifiable Information Physics-Informed Neural Networks Pharmacokinetic/Pharmacodynamic Precision–Recall Curve Prediction model Risk Of Bias Assessment Tool—Artificial Intelligence. Quantitative Polymerase Chain Reaction Quantitative Systems Pharmacology Root Mean Square Error Radiopharmaceutical Therapy Reinforcement Learning Reinforcement Learning-based Dynamic Insulin Titration Root Mean Square Error Receiver Operating Characteristic Real-World Data Real-World Evidence R-squared Software as a Medical Device Severe Acute Respiratory Syndrome Coronavirus 2 SHapley Additive exPlanations Service-Level Agreements Secure Multi-Party Computation Standard Of Care Surface Electromyography Standard Protocol Items: Recommendations for Interventional Trials—Artificial Intelligence Strain Synchrony Index Therapeutic Drug Monitoring Theranostic Digital Twins Telemedical Interventional Management in Heart Failure II Target Registration Error United Nations Educational, Scientific and Cultural Organization Uncertainty Quantification University of Virginia/University of Padova Type 1 Diabetes Positron Emission Tomography Virtual-Heart Ablation-Target Variational Autoencoders Virtual Molecular Tumor Board Virtual Patient Virtual Physiological Human Virtual Reality Virtual Twin Verification and Validation Watt-hour (unit of energy) World Health Organization Explainable AI Three-dimensional 5-Fluorouracil Fifth-Generation Sixth-Generation Authors' contributions The author conceptualized the study, conducted the literature review, collected and analyzed the data, and wrote the manuscript. Funding No funding. Data availability Upon request, data are available. Declarations Competing interests The author declare that she has no conflict of interest. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement Upon request, data are available.