Artificial Intelligence in Tetralogy of Fallot: From Prenatal Diagnosis to Lifelong Management: A Narrative Review Authors: Tiziana Fragasso, Davide Passaro, Alessandra Toscano, Antonio Amodeo, Alberto Eugenio Tozzi, Giorgia Grutter Published: 2025 ABSTRACT Artificial intelligence (AI) is rapidly transforming cardiovascular medicine, with profound implications for congenital heart disease (CHD). Tetralogy of Fallot (ToF), the most common cyanotic disease, requires lifelong surveillance and complex management because of late complications such as pulmonary regurgitation, arrhythmias, and right ventricular dysfunction. This review synthesizes current evidence on AI applications across the continuum of ToF care—from prenatal diagnosis to adulthood follow-up. We examine advances in imaging, perioperative planning, intraoperative monitoring, intensive care, and long-term surveillance, including wearable and implantable technologies. Machine learning (ML), deep learning (DL), and natural language processing (NLP) are revolutionizing diagnostic accuracy, risk stratification, surgical decision-making, and personalized long-term care. The future lies in the integration of multimodal data, including imaging, electronic health records (EHRs), genomic information, and continuous monitoring, to support precision medicine. Challenges remain regarding dataset limitations, interpretability, regulatory standards, and ethical concerns. Nevertheless, ongoing innovation and collaboration between clinicians, engineers, and regulators promise a new era in congenital cardiology. By embedding AI throughout the patient journey, healthcare systems may improve outcomes and quality of life for individuals with ToF. 1. Introduction Tetralogy of Fallot (ToF) accounts for 5–7% of all congenital heart defects and is the most common cyanotic lesion beyond infancy [1]. Advances in surgical techniques and perioperative care have dramatically improved survival, with more than 90% of patients now reaching adulthood [2]. Nonetheless, ToF remains a lifelong clinical challenge, as patients continue to be at risk for both early and late complications, including pulmonary regurgitation, progressive right ventricular (RV) dilatation, arrhythmias, sudden cardiac death, and the need for reinterventions [2]. The annual financial burden of complex congenital heart disease (CHD) is strikingly high, with per-patient costs averaging nearly $50,000, almost half of which is borne directly by families [3]. Importantly, 31% of lifetime costs are sustained within the first five years of life, amounting to an average of $650,000 per child, including about $190,000 in out-of-pocket expenses for families. This early concentration of costs places extraordinary strain on caregivers, who often reduce working hours, with an estimated cumulative productivity loss for parents and patients of roughly one month per year. Compared with more common cardiac conditions, the disproportionate and lifelong nature of these costs underscores the urgent need for greater investment in research, innovation, and institutional support programs to alleviate the economic and human toll of complex CHDs. In parallel, artificial intelligence (AI) is rapidly emerging as a transformative tool in medicine. As of June 2025, the U.S. Food and Drug Administration had cleared 1016 AI-enabled medical algorithms, of which 104 were in cardiology and 777 in radiology—nearly doubling within just two years [4,5]. Machine learning (ML) can identify hidden patterns across multimodal datasets by integrating imaging, hemodynamic, genomic, and wearable-derived data, thereby uncovering clinically relevant associations, predicting patient outcomes, and supporting personalized decision-making that would be difficult to achieve with traditional analytic approaches. Beyond its clinical applications, AI also exerts a significant impact on healthcare economics by optimizing resource allocation, reducing diagnostic and therapeutic inefficiencies, and enabling the earlier detection of complications—factors that could ultimately lower hospitalization rates and long-term costs for public health systems. This review provides a comprehensive synthesis of how AI can support patient care across all stages of ToF management, from the prenatal period to adulthood, supporting diagnosis, treatment, and individualized follow-up. Specifically, we examine:the role of AI in prenatal and perinatal care;applications in perioperative management and early postoperative care;the contribution of AI to long-term follow-up;opportunities arising from data integration and predictive modeling;challenges, limitations, and future perspectives for implementing AI in congenital cardiology. the role of AI in prenatal and perinatal care; the role of AI in prenatal and perinatal care; applications in perioperative management and early postoperative care; applications in perioperative management and early postoperative care; the contribution of AI to long-term follow-up; the contribution of AI to long-term follow-up; opportunities arising from data integration and predictive modeling; opportunities arising from data integration and predictive modeling; challenges, limitations, and future perspectives for implementing AI in congenital cardiology. challenges, limitations, and future perspectives for implementing AI in congenital cardiology. The goal is to provide clinicians and researchers with a practical framework to understand the current state-of-the-art, the potential benefits, and the limitations of AI in ToF, thereby fostering the development of innovative strategies that can be effectively integrated into clinical practice. This review integrates current knowledge from fetal to adult care, offering a roadmap for the role of AI in transforming ToF management. Most of the relevant ToF-related AI studies described in this review are summarized in Table 1, highlighting for each model the input modality, training and validation datasets, clinical outcome, performance metrics, and the presence of an external validation cohort or clinical benchmark. Figure 1 provides a schematic overview of the ToF AI ecosystem, illustrating the data flow from prenatal diagnosis through surgical management to lifelong follow-up. Literature Search Strategy This narrative review is based primarily on a literature search performed in PubMed/MEDLINE. We searched the database from January 2000 to October 2024 using combinations of the following terms: “Tetralogy of Fallot”, “congenital heart disease”, “artificial intelligence”, “machine learning”, “deep learning”, “neural network”, “radiomics”, “clinical decision support”, and “digital twin”. We included original studies and reviews reporting applications of AI to ToF, or to broader congenital heart disease populations including ToF, across the patient journey (prenatal diagnosis, perioperative management, and long-term follow-up). We restricted the search to human studies published in English. In addition, a complementary non-systematic search using Google Scholar was performed to identify further relevant articles, including by screening the reference lists of key papers. 2. Artificial Intelligence in the Prenatal and Perinatal Phase of Tetralogy of Fallot Early detection of CHD, including ToF, is crucial for optimizing perinatal and postnatal outcomes. Prenatal diagnosis facilitates planned delivery at specialized centers, coordination of neonatal interventions, and early parental counseling. However, fetal cardiac imaging remains challenging because it requires substantial operator expertise and is susceptible to technical limitations [23]. AI can improve the accuracy and consistency of cardiac imaging interpretation, particularly in echocardiography and fetal cardiac magnetic resonance imaging (MRI) [24]. Fetal echocardiography remains the cornerstone of prenatal diagnosis for CHD, including ToF, but it is limited by operator variability and diagnostic ambiguity, especially when distinguishing overlapping anomalies such as ToF and isolated ventricular septal defects (VSDs). Convolutional neural networks (CNNs), a type of deep learning (DL) algorithm specifically designed to process visual data, are increasingly used to enhance diagnostic accuracy by automating image acquisition, chamber segmentation, and disease classification. AI-driven systems trained on large datasets improve inter-operator consistency and diagnostic yield. Arnaout et al. (2021) [6] reported that DL-based models achieved 71% sensitivity and 89% specificity for detecting ToF from prenatal ultrasound, notably improving upon traditional diagnostic rates, with real-world sensitivity for the prenatal detection of ToF reported as low as approximately 50% [25]. Similarly, Yu et al. [7] compared four CNN architectures and found that the weakly supervised data augmentation network (WSDAN) outperformed the others, with an AUC of 0.873 in differentiating ToF from VSD. Fetal cardiac MRI, although less widely used than echocardiography because of motion artifacts, lack of ECG gating, and complex acquisition protocols, offers superior spatial resolution and tissue characterization, which is particularly valuable in complex CHD. Uus et al. [8,26] developed AI-based frameworks using CNNs and deformable slice-to-volume registration, significantly enhancing image quality and anatomical interpretability in motion-corrupted datasets and allowing for better visualization of cardiac chambers, vessel alignment, and outflow tract morphology. Among commercially available software solutions, BrightHeart received FDA 510(k) clearance as the first AI software dedicated to prenatal ultrasound of the fetal heart, supporting clinicians in detecting morphological abnormalities suggestive of CHD and reducing inter-observer variability in resource-constrained settings [27]. Other platforms, such as FetalHQ (GE Voluson), also provide rapid AI-driven assessment of fetal heart size, shape, and function in under three minutes; however, variable reproducibility has been observed among operators, influenced by parameter selection, operator experience, and technical factors [28]. 3. Applications in Perioperative Management and Early Postoperative Care In this section, we focus on AI applications during the perioperative window, including preoperative planning, intraoperative guidance, and early postoperative/ICU risk prediction. Medium- and long-term follow-up after discharge are discussed separately in Section 4. 3.1. Preoperative Risk Assessment and Timing/Type of Repair AI applications in perioperative and surgical management span from preoperative planning to intraoperative monitoring and anesthetic management. Surgical repair of ToF typically includes closure of the ventricular septal defect (VSD) and relief of right ventricular outflow tract (RVOT) obstruction, yet the perioperative course remains complex because of anatomical heterogeneity, possible coronary anomalies, and the risk of early complications such as arrhythmias, low cardiac output syndrome, and biventricular dysfunction [1]. AI has been applied to cardiac computed tomography (CT) to improve characterization of right ventricular geometry and pulmonary artery anatomy, enabling more precise delineation of the coronary artery (CA) course, which remains a critical factor in surgical planning and decision-making. Barak-Corren et al. (2025) [9] developed a novel image-derived 3D modeling approach to assess CA proximity to the RVOT in patients undergoing transcatheter pulmonary valve replacement (TPVR) with self-expanding valves, revealing that the anatomical relationship between the RVOT and the CAs was dynamic across the cardiac cycle. In this study, CT-derived segmentations of the RVOT and CAs were created using ML, making 3D segmentation more efficient and enabling processing of multiple frames over time in 4D CT angiography, thus providing a dynamic evaluation of the RVOT. Wang et al. (2021) [10] developed the Structurally Optimized Stochastic Pooling Convolutional Neural Network (SOSPCNN), an advanced DL architecture that combines optimized network design with stochastic pooling to improve accuracy and generalization in medical image analysis. They achieved highly accurate CT-based recognition of ToF, outperforming existing state-of-the-art methods. These tools not only improve anatomical comprehension but also reduce variability between operators and institutions. 3.2. Intraoperative Decision Support In the operating theater, two supervised ML models have been developed to support surgeons during pulmonary valve replacement in ToF patients [11]. Using a novel video-based technique called video Kinematic Evaluation (Vi.Ki.E), Lo Muzio et al. recorded and analyzed the motion of the beating right ventricle (RV)—which cannot be properly evaluated by transesophageal echocardiography—during surgery. They trained k-nearest neighbor and support vector machine classifiers on intraoperative videos of the beating RV in 12 ToF patients undergoing pulmonary valve replacement: videos acquired before valve implantation were labelled ‘unhealthy’, and those acquired after valve replacement ‘healthy’. By extracting motion features from 86 high-speed videos and training these supervised models, they achieved very high accuracy (true-positive rates ≥ 95% and AUC ~0.97–0.99) in distinguishing ‘unhealthy’ pre-replacement from ‘healthy’ post-replacement RV function before chest closure. The models also successfully predicted outcomes in additional patients not included in the training set and may help guide real-time surgical decisions. Such systems may be particularly valuable in minimally invasive or robotic-assisted interventions, where direct visualization is limited. 3.3. Early Postoperative and ICU Risk Prediction The early postoperative phase following surgical repair of ToF is a critical period marked by a heightened risk of hemodynamic instability, ventricular dysfunction, arrhythmias, and respiratory complications. Optimal management during this time requires continuous monitoring, rapid interpretation of complex physiological data, and timely therapeutic interventions to identify patients at risk of rapid deterioration. AI has the potential to support clinicians in risk stratification through predictive modeling. ML algorithms trained on large surgical and imaging datasets can integrate preoperative variables—including laboratory values, echocardiographic data, and clinical scores—to predict adverse outcomes such as arrhythmias, need for prolonged mechanical ventilation, or extracorporeal support. Xi et al. [12] explored risk factors for adverse events and developed ML-based prediction models to forecast the incidence of clinical adverse events after ToF repair. They evaluated five ML algorithms and identified key risk factors such as cardiopulmonary bypass (CPB) time, RVOT pressure gradient, and transannular patch repair. Among all models, logistic regression and Gaussian naïve Bayes performed most consistently, showing good discrimination (i.e., effectively distinguishing patients at higher versus lower risk of postoperative adverse events; see Table 1 for performance metrics), calibration, and clinical utility. The great innovation behind these approaches is that continuous data streams from multiple bedside sources—electrocardiograms, ventilators, infusion pumps, blood gas analyzers, and multiparameter monitors—are already captured within the ICU/EHR ecosystem and can be analyzed in near real-time. This infrastructure provides a natural substrate for electronic clinical decision support systems (CDSSs) and early warning tools that use routinely collected signals to detect subtle signs of impending deterioration, such as changes in cardiac output, rising lactate levels, or evolving arrhythmias, and to trigger timely alerts for the clinical team. In this context, García-Canadilla et al. [29] described CORTEX, a rule-based (not ML-driven) early warning system that uses six age- and physiology-adjusted vital signs (heart rate, respiratory rate, SpO2, systolic and diastolic blood pressure, and temperature) to stratify the risk of clinical deterioration in children after cardiac surgery. In their 1-year pilot study, CORTEX scores increased significantly up to four hours before adverse events, enabling earlier intervention. Its implementation was associated with a 7% reduction in ICU length of stay and an 8% reduction in overall hospital stay, corresponding to approximately €138,000 in total annual hospital savings (not per patient). Importantly, the authors outline the development of an advanced, machine-learning–based upgrade that will integrate more than 50 clinical variables from continuous monitoring, laboratory data, medications, and imaging to enhance predictive performance. 4. The Contribution of Artificial Intelligence to Long-Term Follow-Up of Tetralogy of Fallot This section addresses the medium- and long-term management of repaired TOF, beyond the early postoperative period, including pulmonary regurgitation, RV dilation and dysfunction, arrhythmias, and sudden cardiac death. 4.1. Medium-Term Outcomes After Initial Repair Traditional surveillance protocols—based on rigid, guideline-defined intervals—often fail to reflect interindividual differences in disease progression. AI offers transformative potential to overcome these limitations through dynamic, data-driven monitoring and predictive modeling. One of the most promising frontiers is the shift from fixed follow-up intervals to AI-driven adaptive surveillance. ML models trained on multimodal longitudinal datasets—integrating imaging, electrocardiography, Holter data, biomarkers, and electronic health records (EHRs)—can predict when each patient is at greatest risk for complications. Such models allow clinicians to anticipate adverse events rather than respond reactively. In adults with repaired ToF, Ishikita et al. [13] developed ML models combining clinical and MRI-derived parameters to predict major adverse cardiovascular events (MACE), thereby informing surgical timing and long-term surveillance strategies. In a more recent study, Ishikita et al. [14] compared 5-year MACE prediction in repaired ToF using either a validated ML model or a multidisciplinary team of adult congenital heart disease experts. Model performance was comparable to that of highly experienced physicians (>20 years), and the incorporation of ML improved risk prediction among less experienced clinicians. Importantly, such adaptive surveillance frameworks should be understood as decision support tools rather than autonomous systems. The model’s role is to highlight periods of increased predicted risk, prompting clinicians to consider additional echocardiography, CMR, Holter monitoring, or biomarker testing within predefined protocols. All diagnostic procedures remain under physician prescription, and no validated branching decision tree for adaptive follow-up exists in repaired TOF. Earlier, Samad et al. [15] showed that baseline cardiac MRI features could predict subsequent RV functional deterioration, highlighting the potential for trajectory-based scheduling. More recently, Mayourian et al. [16] applied DL to electrocardiography (ECG) signals, outperforming traditional QRS-duration metrics in predicting mortality. Collectively, these findings emphasize that AI-based dynamic follow-up could replace static, one-size-fits-all models with personalized risk trajectories. Cardiac MRI has so far been the most extensively explored imaging modality in AI research and remains the cornerstone of long-term follow-up in ToF. From acquisition to reconstruction and post-processing, DL algorithms have been shown to improve speed, accuracy, and reproducibility. Most models have been trained and validated using cardiac MRI datasets of structurally normal hearts or cases with acquired cardiac disease and are therefore not well-suited to handle cases with congenital cardiac disease such as ToF. Tilborghs et al. [17] developed and validated a fully automatic segmentation and quantification approach for LV and RV, including RV mass, in patients with repaired TOF with a superior performance in RV quantification compared to commercial software. Govil et al. [18] demonstrated a DL pipeline for automated view classification, slice and phase selection, landmark localization, and myocardial segmentation, enabling end-to-end cardiac shape modeling in repaired ToF. This reduced the time needed to obtain a single cardiac shape model to 5.1 ± 2.8 min, compared with 60–90 min per manual model by an expert analyst. Leiner et al. [24] outlined how ML impacts the full continuum of cardiac MRI—from patient scheduling and image acquisition to diagnostic interpretation and prognostic modeling. Integration of ML within cardiac MRI workflows enables accelerated image reconstruction, artifact correction, and prognostic insight by linking imaging features with clinical and genetic data. Nevertheless, challenges persist, including limited data generalizability across scanners and institutions, data scarcity in rare conditions such as CHD, and the “black-box” nature of DL algorithms [19]. Ongoing work in explainable AI (XAI) seeks to increase transparency and clinician trust, promoting responsible implementation in congenital cardiology. 4.2. Long-Term Surveillance and Adult Congenital Outcomes AI also extends ToF care beyond the hospital setting through continuous physiological surveillance. Smartwatches, patches, and implantable devices generate continuous streams of data on heart rate, rhythm, oxygen saturation, and activity. ML algorithms can analyze these data in real-time to detect subclinical arrhythmias or early hemodynamic compromise, prompting timely telemedicine review and preventing unplanned admissions. Jacquemyn et al. [30] anticipate that decentralized mobile health platforms—integrated with AI-based anomaly detection—will be essential for future ToF management. Recent work demonstrates the clinical feasibility of these technologies: Umer et al. [31] proposed an IoT-based monitoring framework for heart failure in adults. The emerging concept of the digital twin represents the ultimate expression of this approach. A digital twin is a dynamic, virtual replica of a patient that integrates clinical, imaging, electrophysiological, and genomic data to simulate disease trajectories and predict responses to interventions [32]. In ToF, patient-specific digital twins could model RV dyssynchrony, guide pulmonary valve replacement timing, and assess the impact of resynchronization therapy [20]. Computational simulations have already shown promise in predicting valve replacement outcomes, quantifying arrhythmia risk, and visualizing long-term remodeling [33]. From an implementation perspective, building a cardiac digital twin is likely to be highly resource-intensive. Outside medicine, Su et al. recently proposed an activity-based costing model for digital twins in manufacturing, with Ref. [34] showing that development costs are driven by data acquisition, IT infrastructure, and specialized personnel, and can range from a few thousand to almost forty thousand pounds even for relatively simple unit-level twins with limited data streams. In pediatrics, Calcaterra et al. [35] described digital-twin systems as projects whose complexity and costs may be comparable to large scientific endeavors such as the Human Genome Project, underscoring that substantial investment will be required before large-scale deployment. Taken together, these data suggest that cardiac and CHD digital twins will demand significant upfront effort and funding, while formal micro-costing and cost-effectiveness analyses specific to congenital heart disease are still lacking and should be a priority for future research. For AI to achieve clinical maturity, integration into the EHR and workflow is crucial. Automated risk dashboards, structured summaries, and clinician alerts could link AI predictions directly with decision support systems. NLP can extract relevant variables from unstructured clinical notes, operative reports, and imaging narratives, thereby reducing manual data entry and making hidden information available to downstream models. Marelli et al. [21] demonstrated how ML applied to large-scale administrative and clinical datasets can refine the diagnosis of congenital heart disease, highlighting the potential of predictive analytics to uncover patterns that may elude conventional classification. In parallel, Alkan et al. [22] showcased the promise of multi-modal data integration—combining clinical, physiologic, and geometric information—to predict cardiopulmonary exercise outcomes in CHD patients. Together, these studies illustrate how predictive modeling and integrated data streams can not only improve diagnostic accuracy but also anticipate functional trajectories, thereby reinforcing the opportunity for a more proactive and individualized approach to pediatric cardiology care. 5. Challenges, Limitations, and Future Perspectives for Implementing AI in Congenital Cardiology So far, most predictive modeling in congenital cardiology remains at the pre-implementation stage. Despite rapid algorithmic development, only a small fraction of AI tools have reached routine clinical use. Barriers include limited external or prospective validation, poor workflow integration, regulatory friction, unclear return on investment, and slow clinical adoption [36]. Recent systematic reviews in critical care have emphasized that operationalization—the process of moving from proof-of-concept to bedside implementation—requires robust evidence, alignment with clinical workflows, and clear governance structures [36,37]. Expert consensus papers in critical care underscore similar priorities: human-centered design, clinician education, standardized datasets, cross-disciplinary collaboration, and strong institutional governance [38]. Cultural barriers also persist: gaps in clinician AI literacy and, conversely, limited clinical literacy among data scientists; lack of clinician-champions; and workflow disruption that erodes trust. Ethical and legal concerns include bias risks, unclear liability in the clinician–model–vendor triad, uncertain reimbursement, and the need for continuous post-market surveillance. Beyond technical readiness, ethical challenges remain pivotal. Algorithmic decision-making raises questions of accountability, privacy, and patient autonomy. Grote and Berens [39] argue that ensuring fairness and interpretability is as critical as accuracy when deploying AI in healthcare. Clear frameworks for human oversight and continuous post-market monitoring are essential to maintain trust. Liability for AI-supported decisions remains one of the most contentious and least resolved issues in healthcare deployment. In most current implementations, AI tools for ToF (e.g., risk scores, image-segmentation pipelines) are classified as decision support systems rather than autonomous decision-makers; accordingly, physicians and institutions still bear the primary responsibility for clinical decisions and for the way AI outputs are used [39,40]. At the same time, manufacturers and software providers may incur product-liability exposure if systems are defective, poorly updated, or marketed without adequate warnings. Recent European initiatives—including the AI Act and proposals for an AI Liability Directive and a revised Product Liability Directive—seek to clarify the obligations of developers, deployers, and clinicians, but a coherent liability framework specific to medical AI is still evolving, and the dedicated AI Liability Directive has recently been withdrawn, leaving a degree of “liability vacuum” to be filled by national tort law and general product-liability rules [41,42,43,44]. Clinicians understandably fear becoming “liability sinks” for opaque algorithms, which in turn dampens adoption. For congenital cardiology and ToF, transparent model documentation, clear role definitions (AI as a second reader or advisory tool), rigorous validation, and institutional policies on documentation of AI use will be essential to balance innovation with legal and ethical accountability. In the European Union, the Artificial Intelligence Act (Regulation (EU) 2024/1689) [42] introduces a unified, risk-based framework. Most medical imaging and decision support systems will be classified as high-risk, triggering strict requirements for data governance, human oversight, cybersecurity, transparency, and continuous post-market monitoring—aligned with the Medical Device Regulation (MDR) and the In Vitro Diagnostic Regulation (IVDR). These are further reinforced by the NIS2 Directive, the Cyber Resilience Act, the Data Act, and the forthcoming European Health Data Space (EHDS). In the United States, regulation is more fragmented. The Food and Drug Administration (FDA) oversees Software as a Medical Device (SaMD) pathways, including Predetermined Change Control Plans (PCCPs); the Office of the National Coordinator for Health IT (ONC) enforces interoperability rules; and the National Institute of Standards and Technology (NIST) promotes voluntary risk-management frameworks. A pragmatic takeaway for developers is to treat compliance as a design feature—incorporating documentation, traceability, bias monitoring, human-in-the-loop oversight, and incident reporting from the earliest stages [43]. CHD presents unique challenges for AI deployment. Phenotypes are highly heterogeneous and evolve over decades, while traditional evidence-based medicine struggles with small, underpowered cohorts. As rare conditions, CHD populations are inherently limited in size, but the amount of information that can be extracted from each patient is immense, encompassing multimodal and longitudinal data such as echocardiography, cardiac MRI, catheterization, operative notes, wearables, and genomics—often fragmented across EHRs. Targets such as arrhythmia burden or right-ventricular remodeling are dynamic, complicating labeling and calibration. This richness of longitudinal data can, however, empower pattern-recognition models. Privacy-preserving federated learning and synthetic datasets may help overcome single-center limitations, especially for ToF surveillance tasks such as pulmonary valve replacement (PVR) timing and arrhythmia prediction. These approaches can transform data scarcity into an opportunity for multimodal learning across institutions. Cross-cutting barriers are a concern. Data limitations include restricted access to accurately labeled, interoperable datasets; fragmented storage; uncertain ownership; and strict data-protection duties under the European General Data Protection Regulation (GDPR) [40] and the US Health Insurance Portability and Accountability Act (HIPAA) [41]. Beyond the generic issues of data fragmentation and bias, ToF datasets pose specific challenges. First, there is marked class imbalance: even in large institutional archives, the number of ToF patients is small compared with healthy subjects or with more common acquired cardiovascular diseases, which increases the risk of overfitting and unstable performance estimates. Second, “label drift” is intrinsic to lifelong follow-up: surgical techniques, perioperative management, imaging protocols, and even definitions of key outcomes (e.g., “severe” pulmonary regurgitation or “clinically relevant” arrhythmia) have changed over the last decades, making historical labels imperfectly comparable to current standards. Third, sharing pediatric echocardiography and CMR data across centers is particularly difficult because of strict privacy regulations (GDPR/HIPAA), small sample sizes that facilitate re-identification, and the lack of harmonized consent and governance frameworks. Technological challenges relate to “black-box” opacity (difficulty understanding why a model yields a given result), performance drift, weak integration with Picture Archiving and Communication Systems (PACS) and EHRs, and high life-cycle costs for retraining and cybersecurity. Most current ToF-related AI applications still behave as “black-boxes”, providing predictions or segmentations without making their internal reasoning transparent to clinicians. Explainable AI (XAI) techniques can help bridge this gap. For imaging-based models, gradient-weighted class activation maps (Grad-CAM) can be used to highlight the regions of the right ventricular outflow tract, pulmonary valve, or branch pulmonary arteries that drive a given classification (e.g., suspected ToF in fetal or postnatal echocardiography, or the detection of residual lesions on CT/CMR). For perioperative and ICU risk models, Shapley Additive Explanations (SHAP) can quantify how each variable (such as preoperative oxygen saturation, cardiopulmonary bypass time, vasoactive-inotropic score, early arrhythmias, or residual gradients) contributes to the predicted probability of prolonged ICU stay or early complications. Similarly, feature-attribution methods can be applied to long-term risk models of right-ventricular dysfunction, QRS prolongation, or major adverse cardiovascular events, clarifying the relative importance of longitudinal CMR metrics, ECG parameters, exercise capacity, and biomarker trends. Notably, at least three of the studies summarized in this review already incorporate XAI components: Arnaout et al. [6] used saliency maps and Grad-CAM overlays to visualize which regions of fetal cardiac views drove the classification of complex CHD (including ToF); Wang et al. [10] implemented Grad-CAM within their SOSPCNN CT model to highlight the image regions supporting automatic ToF recognition; and Ishikita et al. [13] applied feature-importance analyses in machine-learning models for 5-year MACE prediction in adults with repaired ToF, identifying the most influential clinical and CMR variables. Taken together, these examples show that XAI is technically feasible across the ToF care continuum and, if adopted more widely, could substantially improve clinical trust and model debugging. Performance drift—the gradual change, usually deterioration, in a model’s predictive performance over time in real-world use, even though the model itself has not been modified—typically arises because the underlying data, clinical practice, or patient population has changed. Approaches such as privacy-preserving federated learning and realistically simulated synthetic data may help, but they require robust technical and ethical safeguards and coordinated multinational efforts in congenital cardiology. A short-term roadmap for CHD and ToF should prioritize the creation of federated, multicenter datasets with harmonized labels; pilot use cases such as arrhythmic-risk monitoring and PVR-timing decision support; preregistration of analysis plans; and the implementation of prospective silent trials followed by pragmatic impact studies. Establishing machine-learning operations with version control, bias audits, and post-market surveillance, together with structured human-in-the-loop oversight and training for interdisciplinary teams, will be essential to safely and effectively embed AI into congenital cardiology. Future research should focus on multicenter validation, standardization of AI pipelines, and cost-effectiveness analyses. The involvement of clinicians at every stage of algorithm design is vital to ensure explainability, safety, and real-world relevance. When combined with wearable and imaging data, AI holds the promise of transforming ToF management into a continuous, personalized, and predictive care model, reducing morbidity, optimizing resource use, and improving long-term outcomes. Management of ToF generates highly fragmented data across the life course—ranging from fetal imaging and operative reports to ICU monitoring, longitudinal MRI/echocardiography, and genomic information. AI provides a framework to break down these silos, enabling the construction of continuous, dynamic risk trajectories that recalibrate as new data are incorporated rather than relying on single-episode models. When linked with EHRs and registries, predictive tools can also optimize health-system efficiency by informing the timing of reinterventions, allocating ICU resources, and scheduling advanced imaging during periods of heightened risk. From a research perspective, integrated datasets allow for the discovery of novel phenotypes or endotypes of ToF and can improve the design of clinical trials. Importantly, patient-facing applications—such as mobile platforms offering individualized forecasts and alerts—have the potential to enhance adherence and empower families in shared decision-making. Limitations of This Narrative Review This review is narrative in nature and relies mainly on a PubMed/MEDLINE search complemented by a non-systematic Google search; other bibliographic databases (e.g., Embase, Web of Science, Scopus) and the grey literature were not systematically explored, so some relevant studies may have been missed.