Ai-enabled baseline ECG for staging secondary cardiac damage in severe aortic stenosis undergoing TAVIi: a retrospective study

Background In severe aortic stenosis (AS) treated with transcatheter aortic valve implantation (TAVI), prognosis is strongly influenced by the extent of secondary cardiac damage, but echocardiographic staging is time- and resource-intensive and not universally available. Purpose We aimed to evaluate whether an AI-enhanced 12-lead ECG could reproduce multiparametric cardiac damage staging and support risk stratification along the cardiac damage continuum.​ Methods In this single-centre retrospective cohort, 280 baseline ECGs from severe AS patients scheduled for TAVI at the coordinating center and 88 healthy controls (stage 0) collected from a validated publicly available dataset were analyzed using a deep-learning app-based ECG model, consisting of a modified 16-layer convolutional neural network (CNN) with residual connections, trained on >90,000 ECGs and externally validated, designed to extract quantitative digital biomarkers. Ten clinically selected biomarkers derived from this system were subsequently used as input features to train Random Forest models to predict echocardiographic secondary cardiac damage stages as binary (0 vs 1–4), three-class (0 vs 1–2 vs 3–4), and exploratory five-class (0–4) outcomes; a separate Random Forest model was developed to estimate the risk of post-TAVI permanent pacemaker implantation from baseline AI-ECG features. Performance was assessed by AUC with 95% confidence intervals, and explainability was evaluated using SHapley Additive exPlanations (SHAP) analysis. Results The binary model achieved an AUC of 0.99 (95% CI 0.99–1.00) for discriminating any cardiac damage (1–4) from stage 0. Three-class discrimination between no (0), early-to-intermediate (1–2), and advanced (3–4) damage yielded an AUC of 0.86 (95% CI 0.85–0.88), with SHAP analysis for stages 3–4 highlighting the AI-ECG biomarkers for atrial fibrillation probability and pulmonary oedema as the dominant drivers of advanced-stage assignment. In the exploratory five-class weighted model, AUC was 0.80 (95% CI 0.78–0.82), and the SHAP profile for stage 4 identified left- and right-ventricular dysfunction biomarkers as the principal contributors, consistent with the pathophysiology of end-stage biventricular damage. The dedicated pacemaker model achieved an AUC of 0.68 (95% CI 0.64–0.72). Conclusions An AI-enhanced ECG pipeline applied to a single baseline 12-lead ECG accurately reproduced echocardiographic secondary cardiac damage staging, with strong three-tier stratification that isolates high-risk stages 3–4 and biologically coherent SHAP profiles for advanced damage. These findings support AI-ECG staging as a scalable, low-cost surrogate for multiparametric imaging-based assessment of extravalvular damage in the TAVI work-up and warrant prospective multicentre validation.