Digital Twin models to address long-term treatment toxicities in children and young adults with cancer Authors: Inas Elsayed, Aleksandar Krstic, Luis F. Iglesias-Martínez, Jessica C. Ralston, Walter Kolch Published: 2026 ABSTRACT Although treatment of childhood and young adult cancer has enormously progressed, long-term treatment-related toxicities (LLCTT) prevent survivors from leading a healthy life. Predictive markers are essential for identifying LLCTT early enough to enable personalised therapies that minimise risks. The complexity of LLCTT poses challenges in developing predictive markers using conventional approaches. Here, we provide an overview of how an innovative strategy, Digital Twins, harnesses recent advances in computational modelling to predict and eventually manage treatment toxicities via a personalised approach. We also address the challenges that must be overcome to integrate these models into paediatric cancer care effectively. Introduction Due to early detection and more effective treatment, survival rates of childhood and young adult patients with cancer (CAYAC) have dramatically improved over the last 50 years1,2. With the increased life expectancy of CAYAC survivors, long-term and late cancer treatment-related toxicities (LLCTT) emerge as significant challenges. A considerable proportion of survivors experience one or more LLCTT, with hazard ratios (HR) of 4.2 in young adults and 5.6 in childhood cancers, contributing to reduced life expectancy (mortality HR = 5.9 in young adults and 6.2 in childhood cancers), chronic morbidity, and diminished quality of life1,3–5. This highlights the need for predictive biomarkers to guide treatment decisions. These biomarkers should identify patients at higher risk for LLCTT, allowing for tailored treatments that minimise adverse effects while preserving positive outcomes. The incidence of LLCTT varies depending on the type of cancer, treatment protocols, and individual patient factors6. Multiple studies have investigated treatment-related chronic conditions to identify predictive markers of these conditions. Several modifiable risk factors have been identified, such as a sedentary lifestyle and obesity, that increase the incidence and severity of LLCTT7,8. However, individual variations in LLCTT suggest the contribution of genetic factors that, if dissected, can provide insights into the mechanisms of these toxicities and identify markers for predicting patients at high risk of developing LLCTT-related morbidities. Numerous genetic variants have been linked to a higher risk of LLCTT in candidate gene studies4,9,10. This was further enhanced by comprehensive OMICs approaches, primarily genome-wide association studies (GWAS) and next-generation sequencing (NGS)11–13. These studies identified genetic variants and gene expression patterns associated with toxicities (Supplementary Table 1). Nevertheless, this has not led to decisive progress in the clinical care of childhood cancer patients, as only a few genetic tests have been successfully validated and implemented in clinical practice for LLCTT prediction14. The variety of CAYAC subtypes and treatment modalities markedly confounds toxicity outcomes, making it difficult to compare and standardise findings across different studies. Moreover, the multifactorial nature of late toxicities, involving genetic, environmental, and treatment-related factors, makes it challenging to identify genetic contributions with certainty9,15. This is further complicated by the genetic diversity among CAYAC survivors from different genetic ancestry groups, where the variability of allele frequencies hampers the identification of broadly applicable predictive markers. This challenge is particularly acute for rare variants that may be significant in one population but too rare to be identified in another16. Thus, classic reductive and static models used to study such toxicities and dissect individual markers have enjoyed limited success17. By relying on isolated data points, these frameworks struggle to capture longitudinal events driven by the interplay between a survivor’s genetic predisposition and evolving variables, such as ageing, development, health condition, and lifestyle18–21. There is, therefore, a critical need for integrated, dynamic models capable of synthesising genetic, molecular, clinical, and other risk factors into robust predictive tools. A promising approach in this context is Digital Twin models (DTs). DTs enhance the understanding of CAYAC LLCTT by integrating multimodal data DTs are virtual replicas of physical entities or processes that enable predictive simulations of the behaviour of these physical systems (Fig. 1 and Text Box 1). When applied to precision medicine, DTs can use a variety of molecular and clinical patient data as inputs, applying advanced computational models and simulations to predict outcomes and optimise treatments. In the context of identifying markers for treatment-related toxicities in CAYAC, the DT framework enhances the utility of classic computational and statistical models. A DT acts as an integrative framework that leverages its predictive power from a bidirectional information flow between the physical patient and their digital counterpart22. A primary strength of the DT frameworks is the scalability and flexibility in modelling, which allows for the simulation of LLCTT at various levels of granularity. Given the inherent heterogeneity of LLCTT phenotypes, this multi-scale approach enables a DT to balance model complexity with data constraints, ensuring an effective and tailored prediction of outcomes23. Fortunately, we are currently experiencing an increasing availability of CAYAC survivorship data. Several survivorship programs have implemented follow-up procedures, including self-reporting questionnaires, survivorship registries, and periodic clinical assessments, to generate data on LLCTT incidence, severity, and outcomes (Supplementary Table 2). Electronic health records also offer structured data resources regarding patient treatment, clinical examination, and laboratory tests. The growing use of comprehensive genomic analysis technologies generates rapidly expanding data on potential genetic, epigenetic, and gene expression-related risk factors for LLCTT24. Continued innovation, such as the Internet of Things, wearables and other connected technologies, could further enhance follow-up and data collection from cancer survivors25. However, integrating such vast amounts of heterogeneous data from various sources to identify meaningful LLCTT associations remains a significant challenge. The emergence of Artificial Intelligence (AI) and machine learning (ML) algorithms capable of integrative multimodal data analysis offers a transformative solution, particularly through their utility within DT frameworks that can simulate and predict complex biological outcomes. However, the successful deployment of these models in the DT framework requires rigorous harmonisation of input data and continuous system updates, which must include a systematic evaluation of how these updates influence specific model parameters and predictive stability26.Fig. 1Integrating multimodal data for the predictive analysis of CAYAC LLCTT.DTs platforms allow for multimodal data integration and analysis by in silico simulations to study the mechanisms of LLCTT development and identify potential predictive biomarkers. CAYAC: Childhood and young adult patients with cancer, LLCTT: long-term and late cancer treatment-related toxicities, UC: Uncertainty Quantification. Integrating multimodal data for the predictive analysis of CAYAC LLCTT. DTs platforms allow for multimodal data integration and analysis by in silico simulations to study the mechanisms of LLCTT development and identify potential predictive biomarkers. CAYAC: Childhood and young adult patients with cancer, LLCTT: long-term and late cancer treatment-related toxicities, UC: Uncertainty Quantification. Several solutions for data management for DT models are available, including architectures specifically designed for medical DTs27,28. Currently, there is no consensus solution, but it is important to utilise standardised formats and ontologies in order to ensure comparability and interoperability of LLCTT data across different studies and institutions28–30. Further, such platforms should comprise workflows for data preprocessing, including data cleaning, imputation, and normalisation, as well as their transformation into structured formats useful for data integration31. The use of feature selection methods enables the retention of relevant features and the removal of irrelevant features that cause noise or bias32–34. The automation of these workflows facilitates updates, ensuring that the integrated datasets reflect the latest findings and observations25,28,30,35. Masison et al.28 developed a resilient modular DT model capable of robustly modelling these interacting and evolving processes. Masison’s model comprises separate modules for the different dynamic processes of the model, which allows updating and scaling of the modules without the need for redesigning the model with each update. Such flexible data management platforms seem ideal for LLCTT DT models, which need to be regularly updated over long observation periods. A Digital Twin (DT) is a virtual representation of a physical entity, a process or system that accurately mirrors the behaviour and functionality of its real-world counterpart. First used in engineering, DTs integrate data and updates from the physical entity about internal and external conditions to simulate adaptations to changes in these conditions in the virtual DT. Such updates are multi-sourced data flows that supply the DT model with information pertinent to the outcomes that are to be simulated. The simulations allow the analysis of mechanisms underlying past behaviour, the prediction of future behaviour, and the design and optimisation of interventions to control future behaviour. In healthcare, DTs can model cells, organs, patients, or entire healthcare systems, facilitating the development of personalised treatments and improved patient outcomes22. DTs can be classified as component twins (modelling a single part of a system), asset twins (modelling multiple parts of a system), system twins (modelling the whole system), or process twins (focusing on how parts of a system work together)22. From mechanistic to data-driven models: Enhancing DTs for personalised LLCTT Prediction The pathogenesis of LLCTT is a complex, multi-step process. Anticancer treatment can cause subclinical damage that triggers pathological molecular and cellular processes, such as immunologic and metabolic reprogramming, which can propagate to clinical manifestations over time36–38. Mechanistic models that simulate these system processes in DT frameworks can unveil the trajectories of LLCTT development at various stages and scales. For example, Strocchi et al.39 described several spatially and temporally scalable models that simulate the complex mechanisms of long-term effects, focusing on the multi-scale processes implicated in cardiovascular diseases caused by diabetes. These models encompass metabolism, electrophysiology, calcium flux at the cellular level, blood flow within the circulatory system, and the effects of drugs on these systems. DTs can integrate such mechanistic models to simulate LLCTT’s pathologic mechanisms at multiple levels, continually refining predictions through iterative updates with real-world patient data39. Additionally, quantitative systems pharmacology (QSP) models are commonly used mechanistic models, which have demonstrated effectiveness in modelling treatment outcomes and drug toxicities in cancer40,41. Diakonova et al. successfully used a QSP model to construct DTs that simulate the responses of head and neck cancer patients to different drug combinations. The model is composed of several sub-modules that simulate the cellular composition of the tumour microenvironment, interactions between the different cell types, and the cytokine profiles in response to various chemo- and immunotherapies, along with the pharmacokinetics of these therapies40. The DT model could potentially describe the effects of various treatments and predict clinical outcomes. Applied to LLCTT studies, such models could create detailed, dynamic simulations of a child’s response to treatment over time, capturing any changes in the system caused by cumulative treatment effects, changes in developmental stage, or lifestyle. Such models are highly advantageous in studying high-risk patients’ response to anticancer therapy. Those patients are usually excluded from clinical trials and hence understudied. Mechanistic DTs like the QSP-based model of Sang et al. successfully generated DT models of patients at risk of cardiotoxicity from chemotherapy and targeted therapies based on data obtained from cardiomyocytes grown and treated in vitro and in vivo models41. This work opens the perspective to use the wealth of in vitro drug toxicity data to parameterise DT models of LLCTT. However, as these models are typically knowledge-based, their utility for DTs is constrained by the limited understanding of LLCTT biological processes and causative factors underlying LLCTT (epistemic uncertainty), and LLCTT heterogeneity (aleatoric uncertainty)42. The advent of genetic studies at scale identified genetic features and polygenic risk scores (PRS) associated with LLCTT (Supplementary Table 1). Unfortunately, many associations are weak, inconsistent, or require independent validation11,43. Thus, not only more data but also better analysis methods are needed. Statistical and ML models in the DT framework can be beneficial, as they provide a rigorous framework for uncertainty quantification and residual error estimation26. One approach is to test PRS derived from the general population in CAYAC patients. This helps distinguish true from spurious associations due to increased sample sizes and reduction of ethnic biases in the general population. This strategy was effective in predicting subsequent malignant neoplasms in CAYAC survivors, as the potential common aetiology with other cancers justified the use of general PRS in the context of CAYAC LLCTT44,45. Also, advanced statistical modelling revealed that genetic and treatment risks combined to predict the occurrence of subsequent neoplasms caused by chemo- and radiation therapy. Interestingly, lifestyle factors, such as obesity, smoking, and diet, only had a small effect10. In particular, mutations in DNA repair genes were linked to the incidence of subsequent neoplasms46. Integrating PRS and clinical data in a novel pharmaco-radiogenomic PRS, which accounts for interactions between genes and treatments, could better quantify the probabilistic risk of survivors developing late pulmonary complications than current clinical predictors47. Using both descriptive statistical models and multivariable subdistribution hazards models suggested that radiation therapy to the chest increases the risk of coronary heart disease in later life48. A wide range of ML models is available, and we are now witnessing their deployment for long-term predictions26,49. The transition toward DTs necessitates ML models that not only predict outcomes but also evaluate the stability of those predictions. An Elastic Net model, i.e., a linear regression ML model with regularisation to avoid overfitting, was used to identify biochemical and molecular markers of premature ageing in CAYAC survivors, a systemic late effect increasingly recognised as LLCTT36. Ensemble learning methods enable the accurate prediction of LLCTT while quantifying uncertainty, moving further towards DT modelling50–52. For instance, Random Forest (RF) modelling could match new patients to previous patients based on abdominal shape and organ positions to determine the optimal radiation dose for new CAYA patients with kidney cancer53. When adapted for 30-year survival analysis and the risk of subsequent neoplasms, RFs provide a distribution of potential outcomes, yielding highly accurate, personalised predictions that account for long-term variability54. Combining ML models with complementary models, such as graph analysis, allows us to analyse complex and difficult-to-track LLCTT, such as cognitive defects55. Diffusion tensor imaging, a magnetic resonance imaging (MRI) technique that visualises the diffusion of water molecules along anatomical tracks, revealed altered connections between brain areas of Acute Lymphoblastic Leukaemia (ALL) survivors. Using RFs to combine Graph Theory analysis of the connection network with demographic and medical features (gender, racial minority status, maternal education level, age at diagnosis, treatment dose, and time since treatment end) produced a model that accurately identified ALL survivors with cognitive defects55. The accuracy of such models could be further improved by adding relevant clinical features, such as behavioural phenotypes, and using ‘stability selection’ to find features that are resistant to perturbations of the training data and hence carry robust information56. Applied to DT frameworks, such models will ensure the information used for the DT is robust and the associated uncertainty is minimised. Further enhancements to the models come from deep learning AI methods, especially their application to advanced image analysis. A neural network method, Kohonen Self-Organising Maps (KSOM), was used to distinguish leukoencephalopathies caused by chemotherapy from normal brain maturation in ALL survivors57. KSOM arrange their ‘neurons’ in a two-dimensional grid that reflects the topology of the input space, allowing a visual interpretation of the modelling results, which is ideal for image analysis. ALL survivors were examined 6, 14, 63, and 135 weeks after methotrexate treatment, with the model correctly classifying ~90% of children with leukoencephalopathy57. Using deep learning to analyse standard of care computer tomography images from CAYAs with lymphoma for body tissue composition showed that patients with more adipose tissue and less skeletal muscle were more likely to experience LLCTT58. This study was retrospective, but the three-year LLCTT outcomes suggested that this method could be employed to identify patients at risk for LLCTT. In the realm of genomics, generative adversarial approaches—such as BRI-Net—address data scarcity by generating synthetic data to bridge gaps in clinical datasets. BRI-Net integrates existing biological knowledge into the architecture of neural networks and trains these networks by a generative adversarial approach59. This involves a generator network that generates synthetic data and an adversarial discriminator network that tries to distinguish the synthetic data from the real data. The training completes when the discriminator cannot distinguish synthetic data from the real data60. Using genomic data from 362 CAYAs treated with cisplatin, BRI-Net identified two SNPs in the CERS6 and TLR4 genes that are strongly associated with cisplatin-induced ototoxicity59. Hybrid frameworks that leverage mechanistic models in conjunction with data-driven statistical/ML/AI algorithms are gaining momentum in oncology research and may provide significant value in addressing LCCTT61,62. By learning from new data, AI models help inform mechanistic models through parameter identification, uncertainty estimation, and model calibration, leading to improved accuracy and predictive capabilities63. The bidirectional flow of information between the biological system (a child with cancer) and the DT enables continuous updates with new data, facilitating ongoing adjustments and refinements in predictions and interventions (Fig. 2)33. Hence, the resulting DTs could emulate the evolution of patients’ responses to treatment and associated outcomes.Fig. 2DTs are predictive personalised models.DTs continuously utilise real-time patient data to optimise the model to mirror the patient’s status and provide real-time individualised predictions, allowing for optimising therapeutic adjustments for each patient. CAYAC: Childhood and young adult patients with cancer, ML: Machine learning, AI: Artificial intelligence, DT: Digital twin. DTs are predictive personalised models. DTs continuously utilise real-time patient data to optimise the model to mirror the patient’s status and provide real-time individualised predictions, allowing for optimising therapeutic adjustments for each patient. CAYAC: Childhood and young adult patients with cancer, ML: Machine learning, AI: Artificial intelligence, DT: Digital twin. While the models described above do not yet represent full DT models, they represent critical foundational stages. By adopting a ‘fit-for-purpose’ approach to model selection and establishing a standardised framework for verification -covering both model performance and bidirectional data flow-, validation, and uncertainty quantification, these models can successfully transition into the realm of biomedical DTs. Challenges and new developments One of the primary challenges in implementing DTs for CAYAC is the limited availability of data for model training and validation. Unfortunately, preclinical animal model systems often fail to predict human anticancer drug toxicities reliably64. The relative rarity of CAYAC, combined with the logistic constraints of extended follow-up, restricts the sample sizes of most genetic studies and clinical trials involving CAYAC survivors. Large-scale survivorship studies are challenging as longitudinal data collection and follow-up require continuing patient compliance, intensive financial resources and organisational efforts. Fortunately, several large-scale survivorship studies, such as the St. Jude Lifetime Cohort Study (St. Jude LIFE) and the Childhood Cancer Survivor Study24,65, are expanding and providing Genomic plus long-term follow-up medical data for developing predictive models (Supplementary Table2). Nonetheless, the representation of survivors’ diversity, including race, ethnicity, and socioeconomic groups, remains suboptimal. The increasing availability of data is more than matched by rapid advances in computational methods. New modular software architectures make DT models scalable and adaptable to accept varying input data and updates28. By generating simulated input data, researchers can enhance diversity and representation without the need to extensively increase the number of physical participants66,67. Synthetic DTs of patients expand the possibilities of virtual experiments and clinical trials that otherwise would not be achievable in real life due to technical and recruitment limitations66,68 (Fig. 3). A major initiative is the ‘Virtual Child’, which uses knowledge graphs to integrate multi-omics, clinical and imaging data from histopathology and macro-morphologic imaging69. The Virtual Child focuses on brain tumours but aims to include other CAYA cancers in the future. However, while generating simulated input data can resolve privacy and limited datasets issues, this approach introduces specific risks regarding algorithmic bias. Synthetic data is inherently limited by the biases present in its training source. If the generative model lacks sufficient grounding in marginalised groups, it may produce hallucinated diversity that fails to capture the true variance of a physical population. Thus, it is vital to implement rigorous validation frameworks to ensure simulated data clinical safety and representational fidelity70.Fig. 3DTs enable virtual LLCTT experiments and clinical trials.DTs can help resolve scale, inclusivity, and reproducibility challenges in LLCTT research by simulating different conditions in experiments and clinical trials. DT: Digital twin, LLCTT: long-term and late cancer treatment-related toxicities. DTs enable virtual LLCTT experiments and clinical trials. DTs can help resolve scale, inclusivity, and reproducibility challenges in LLCTT research by simulating different conditions in experiments and clinical trials. DT: Digital twin, LLCTT: long-term and late cancer treatment-related toxicities. The rapid improvements of large language models also offer unprecedented possibilities to include the wealth of existing medical knowledge with an intuitive interface for interrogating the models using plain language queries and answers, although efforts to standardise these methods are required71. An exciting development are the Foundation Models, which are trained on large amounts of general data. They can find hidden patterns in the data, and when applied to target data, perform remarkably well with little further training, even with limited sample sizes. For example, Foundation Models for histopathology accurately diagnosed different cancer types and stages72. It is crucial to acknowledge the inherent limitations of such AI methods. On their own, they cannot explain the patterns or correlations observed within the data unless combined with mechanism-based models62,73. Combining statistical models and AI/ML that can recognise and predict data patterns and behaviours, with mechanistic models that can deliver tractable explanations for the variations caused by the system’s variable interactions, like cellular and molecular interactions, is a recommended strategy in the context of DTs28. Illustrating the practical application of these hybrid approaches that combine AI/ML with mechanistic models is the cSTAR (cell state transition assessment and regulation) method. cSTAR uses AI/ML to classify cell states and extracts mechanistic models that can be populated with patient-specific data74,75. Predictions are made by the mechanistic model, which delivers fully tractable and dynamic simulations. The possibility to explicitly model changes in response dynamics over time makes these models ideal to construct DT that can simulate acute and, if longitudinal data are available, also long-term treatment toxicities and differential manifestations in individual patients. Verification and validation are fundamental to ensuring the safety and efficacy of digital twins (DTs). Verification confirms that a model is correctly implemented according to its mathematical and computational specifications, while validation assesses whether the DT accurately represents the real-world biological system it intends to simulate42. The dynamic nature of DTs demands a verification framework that considers any variation in input data quality or model parameters over time that might affect the model’s integrity. Quantification of uncertainty, as part of model verification, is an equally critical component of DT trustworthiness. Uncertainty can arise from data acquisition, model selection, or the model’s parameters. To balance interpretability with computational efficiency, strategies such as surrogate modelling, scientific machine learning, and multi-fidelity methods can be implemented. However, as uncertainty is inherent in biological modelling, a strategic trade-off is often required: intervening early with limited data and higher uncertainty versus intervening later when more data reduces uncertainty but potentially diminishes the clinical impact42. Validation of DTs remains challenging due to the absence of standard evaluation metrics, the lack of ground-truth data for prospective predictions, and the context variation between DT development and clinical application. Setting a framework for model validation and revalidation -as needed- is essential for DTs to ensure models’ reproducibility over time42,62. Moreover, external validation of DTs is necessary to ensure the generalisability of results, as these models can be susceptible to bias or overfitting. In contexts such as LLCTT, where patient heterogeneity is high and rare conditions may limit the availability of independent data points for validation beyond the initial calibration set, N-of-1 personalised trials—which randomise treatment periods for individual survivors—offer a viable path for validation42. Finally, beyond standard metrics, a clinical utility threshold should be integrated into the model evaluation to determine whether DTs genuinely improve decision-making and patient outcomes compared to standard-of-care or established models76. Implementation of DTs in the clinical care pathway of CAYAC patients Integrating DTs into the routine clinical care pathway holds tremendous potential for minimising LLCTT. With the increasing use of electronic health records and Internet of Things applications, implementing DTs in medical care is now more plausible than ever, despite remaining challenges. A main challenge is ensuring the quality, reliability, and interoperability of data obtained from such electronic health records77,78. In addition, DT models must be rigorously validated before introduction into clinical settings. This includes testing in clinical studies, compliance with specifications and legal requirements for clinical decision support systems, and approval by regulatory authorities79. In addition to technical and clinical validation, health economic assessments are also needed. Integrating DTs into the existing clinical workflows requires substantial investment in staff training as well as technology and software. This includes the setup and maintenance of advanced computational resources, robust cybersecurity measures, and solutions for data storage, management and governance79. Moreover, the integration of DTs in clinical practice requires multidisciplinary collaboration across multiple departments, including clinicians, bioinformaticians, IT technicians, and policymakers. Coordinating efforts among these diverse fields can be tedious, requiring training, effective communication, and interdisciplinary understanding79. DTs also raise new types of ethical issues and privacy concerns, especially the use of ML/AI and the personalised models, which will require new regulatory guidelines. Ensuring equitable access to personalised treatments across various socioeconomic and racial groups is another significant task80. However, DTs also offer the chance to democratize medicine, as these models can operate remotely and, when established at scale, allow broad and equitable access across socio-economic barriers. Conclusion and Outlook In conclusion, DT models represent a transformative approach to studying and predicting CAYAC LLCTT. While the full integration of DTs into routine pediatric oncology and CAYAC healthcare remains a long-term goal, the next five years will be a critical period for establishing prospective pilot studies and clinical validation frameworks. This progression is underpinned by the recent expansion of the CAYAC survivorship data and the availability of large-scale genetic and phenotypic datasets (Supplementary Table 2), necessary for high-fidelity modelling. Furthermore, efforts to standardise longitudinal clinical and molecular data collection and sharing -exemplified by the National Cancer Institute’s Childhood Cancer Data Initiative (CCDI)81- combined with recent breakthroughs in computational modelling, have significantly advanced the field. The assignment of optimal drug therapies for paediatric cancers through computational models is already at a stage where it has clinical utility. A recent study showed that digital drug assignment identified appropriate drugs for 88% of high-risk paediatric cancer patients. Seventy-eight per cent of the suggested drugs were subsequently approved by a Molecular Tumour Board and used in clinic82. In the same vein, a retrospective study of 1,151 CAYACs focusing on 10 genes and 28 drugs showed that pharmacogenetic screening based on Genome sequencing could have improved treatment efficacy and reduced side effects in 16% of patients83. Consequently, we anticipate that DT models for LLCTT will transition from retrospective research tools toward investigational decision-support systems in specialised centres. This transition represents an essential phase of rigorous clinical evaluation, paving the way for eventual evidence-based adoption. The potential applications of DTs extend to the identification of biomarkers for early detection of LLCTT, optimising treatment regimens, and improving the long-term survivorship outcomes of CAYAC patients. Leveraging real-world data and advanced computational techniques, DTs offer a personalised framework for simulating patient responses to cancer therapies over time, enabling clinicians to tailor interventions based on individual risk profiles22. Looking forward, integrating DT models in CAYAC clinical care requires exploring the avenues of related economic, logistic, and ethical issues. Investing in technology infrastructure and data-sharing endeavours is crucial while fostering interdisciplinary collaboration among researchers, clinicians, and data scientists to develop and implement robust DT platforms. Additionally, establishing ethical and legal regulations will be essential to fully realise the potential of DTs in clinical practice. Supplementary information Supplementary Information Supplementary Information