Generalisable artificial intelligence ECG trained on public data for outcome prediction after transcatheter aortic valve replacement

Background Artificial intelligence ECG (AI-ECG) models can predict cardiovascular outcomes, but their clinical adoption is limited by restricted access to training data and uncertain generalisability. We developed and externally validated a generalisable AI-ECG model trained exclusively on publicly available data to predict outcomes after transcatheter aortic valve replacement (TAVR). Methods A transformer-based AI-ECG model was trained on 341 151 publicly available ECGs with mortality labels. The model was externally validated in a prospective cohort of 439 patients undergoing TAVR and additionally tested in a surgical aortic valve replacement cohort. Model performance for predicting 30-day and 1-year mortality was assessed using area under the receiver operating characteristic curve (AUC) and associations with clinical outcomes were evaluated using multivariable regression. Results In the TAVR cohort, a single pre-procedural ECG predicted 30-day mortality with an AUC of 0.85 (95% CI 0.78 to 0.92) and 1-year mortality with an AUC of 0.74 (95% CI 0.67 to 0.80). The derived AI-ECG risk score was independently associated with 1-year mortality (adjusted OR 1.70), major adverse cardiac events and major renal events. Predictive performance was consistent in an independent surgical cohort. High-risk ECG features included atrial fibrillation, prolonged QT interval, ST-T abnormalities and low voltage. Conclusions An AI-ECG model trained solely on publicly available data provides accurate and generalisable prediction of outcomes after TAVR using a single ECG. This approach demonstrates the feasibility of transparent, shareable AI models for cardiovascular risk stratification and supports a general-to-specific paradigm for clinical deployment.