Machine learning in mitral valve repair: promise, precision, and the challenge of clinical translation Authors: Eline Kho, Maurice A. P. Oudeman, Henning Hermanns, Denise P. Veelo Published: 2026 In a recent issue of the Journal of Thoracic and Cardiovascular Surgery, Malik and colleagues present a machine learning (ML)-based model to predict long-term durability after mitral valve repair, comparing a random survival forest approach with traditional Cox proportional hazards modeling (1). The authors are to be congratulated for applying ML to an important and clinically relevant question, and for achieving excellent surgical results, reflected by a remarkably low event rate and near-complete elimination of significant residual mitral regurgitation (MR) at the end of surgery. At the same time, in our view, several aspects warrant cautious interpretation. Mitral valve repair failure is widely recognized as a multifactorial process, with residual MR ≥2+ consistently emerging as a relevant predictor of poor durability, alongside age, anatomical factors (e.g., coaptation length, annular dimensions, chordal integrity), ventricular function, transmitral gradients, pulmonary pressures, and procedural complexity (2-4). Additionally, there is growing recognition that functional MR itself comprises heterogeneous phenotypes, including atrial and mixed forms, with differing pathophysiology and therapeutic implications (5). Existing literature, including studies on early versus late failure and alternative predictive models, underscores this multidimensional nature (6). In contrast, the present analysis appears to reduce prediction largely to coaptation length and transmitral gradient. This discrepancy may reflect limitations in capturing established determinants, but could also be explained by the relative homogeneity of a highly selected single-center, single-surgeon cohort. Indeed, such excellent and consistent surgical performance, while commendable, may paradoxically limit variability and thereby constrain model generalizability, suggesting that the model may primarily reflect surgeon-specific outcomes rather than a broadly applicable algorithm. This interpretation is further supported by other ML-based analyses using similar methodologies that have identified different sets of predictive factors, underscoring the potential dependence of such models on local practice patterns and cohort characteristics rather than universal determinants (7). From a clinical perspective, it is also important to consider that key input variables, particularly intraoperative valve performance parameters such as transmitral gradient, are dynamic and strongly influenced by loading conditions, anesthetic management, ventilation, and heart rate (8). While dynamic intraoperative variables are undoubtedly clinically valuable and may improve predictive performance, their interpretation requires careful consideration because they are highly dependent on loading conditions and anesthetic management. Consequently, prediction models using these variables should acknowledge this inherent variability to avoid a sense of false precision. From model development to implementation Beyond these clinical considerations, the study by Malik et al. also reflects a broader pattern in the field: many ML algorithms are developed and internally validated, yet relatively few progress to external validation, demonstration of clinical effectiveness, and ultimately large-scale (9,10). The lack of external validation might partly be the result of a logistical challenge, as acquiring high-quality datasets from independent institutions can be difficult. However, a more important barrier lies within the nature of the data the model is trained on. Clinical data are often heterogeneous across, and even within, institutions. Here, variations arise due to different electronic health record systems, but also differences in terminology, as even within a single institution, variations in the annotations and definitions, and even language, have been documented (11). The deviations are likely amplified in multi-center settings, posing real challenges for external validation. Instead of external validation, temporal validation is frequently reported, where data is collected at the same site as the training data, but from a different time period. However, a good temporal validation outcome gives no guarantee of transportability across institutions. Second, an extensive increase in published artificial intelligence-based predictive or detection models shows the rapid development within this field. However, while the methodological innovation is high, alignment with a clearly defined clinical need is crucial. Even models with excellent performance face implementation issues if the problem they address lacks a strong clinical use case, and may eventually result in a lack of trust (12). While this issue does not apply to the work by Malik and colleagues, as they present a model to improve current practice, it underscores the importance of need-driven model development. Third, model implementation is strongly affected by access to knowledge and information, the data infrastructure, and organizational culture (13). The concept of a ‘basic’ model that can be adapted to the local need and context is appealing and also in line with the trend towards personalized health care. However, fundamental issues need to be dealt with, concerning the responsibility for the basic model, but also the ownership of the tailored model. It might be unrealistic to expect clinicians to master the technical aspects, in addition to their clinical responsibilities, to tailor and maintain these models, as after implementation the model needs to be kept up to date. Instead, we might need dedicated health care professionals with a technical background who bridge the clinical and technical domains alongside the clinicians. In addition, these prediction models are particularly vulnerable facing alternations affecting the model input parameters, like changes in surgical procedures, diagnostics or target patients. Explainability versus effectiveness When publishing ML (or other prediction/detection) models, there are a lot of different outcomes to show. Which one to show depends on the type of data used for the model, what part in the model development is being tested, the balance of data, or the clinical purpose. Several different performance measures are cited in literature, where every performance has its own characteristics (14). When the class distribution of a dataset is imbalanced, reporting accuracy can be a misleading evaluation metric, as it illustrates the overall proportion of correct predictions, not considering the imbalance. Whereas calibration might be more insightful for external validation. A decision curve might be helpful in showing the net benefit of its clinical utility. The work by Malik and colleagues also highlights an important topic concerning the balance between the explainability of a model versus its effectiveness. Highly complex models may achieve excellent performance but lose explainability and function as a ‘black box’, limiting transparency in the decision-making processes. However, this raises a practical question: to what extent does a clinician need to understand the exact mechanisms behind a model to still be able to responsibly use it? There is a difference in understanding how to use a model and understanding how it works. Users should be able to interpret the model outcomes and understand its limitations, whereas understanding underlying algorithms might be embedded within the domain of technical experts. The evolution of artificial intelligence in healthcare is a continuous process, where with every new step, new discussions surface. However, as these technically more advanced models might help us face the increasing work pressure in healthcare, we welcome these discussions. Beyond improving efficiency, these models have the potential to advance personalized medicine, where abnormalities might be detected earlier, timely intervention can be given before the onset of irreversible remodeling, and complex patterns, invisible to human observers, can be detected. Conclusions Taken together, this study represents an innovative first step toward personalized prediction in mitral valve repair. However, careful clinical contextualization, methodological expansion, and rigorous validation will be essential to translate such promising approaches into robust and widely applicable tools for routine practice. Supplementary The article’s supplementary files as Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1322/prf