Abstract
Background: Echocardiography is widely available and provides functional information about the heart, but is rarely used systematically for infarct localization in myocardial infarction (MI). Regional longitudinal strain (RLS) carries territory-specific information that may help identify ischemic regions and support early triage to coronary intervention, particularly in patients without clear STEMI criteria. Classical machine learning methods such as tree-based classifiers typically require manual feature engineering and task-specific tuning, and they therefore often generalize poorly to new cohorts. Tabular foundation models are a novel alternative within deep learning, where a single model is trained once on a broad range of data, then adapts to a new dataset at prediction time, without fine-tuning for downstream tasks.
Aim: To evaluate whether a tabular foundation model can predict CA occlusion in an MI cohort, and to compare its performance with established tree-based machine learning methods.
Methods: We included 1404 patients (mean age 74±12 years, 76% male) with MI and prior coronary revascularization. Inputs included age, sex, and echocardiographic parameters derived from an automated algorithm: RLS per segment, ejection fraction, mitral annular plane systolic excursion, and mechanical dispersion. For each of the CAs (left anterior descending artery, LAD; circumflex artery, CX; right coronary artery, RCA), a separate binary classifier predicted total occlusion. We compared TabPFN, a foundation model without any dataset-specific tuning, against random forest and XGBoost.
Results: TabPFN outperformed both reference methods across all three coronary territories. Mean AUC was 0.79±0.02 overall, with the strongest performance for LAD occlusion (0.84±0.04), followed by RCA (0.78±0.04) and CX (0.76±0.03).
Conclusion: A tabular foundation model, applied without any task-specific tuning, matched or exceeded tree-based machine learning methods optimized for the task. This approach may reduce the development burden for clinical decision-support tools based on structured clinical data, though validation in larger, multi-centre cohorts is needed.