Abstract 4369502: Identifying optimum ECG features to predict sudden cardiac arrest at varying time points before the event

Abstract

Background: Prediction of sudden cardiac arrest (SCA) from the ECG is of great importance, but is complicated by the uncertainty introduced by dynamic remodeling of the ECG over time. Traditional machine learning models trained on ECG snapshots have not captured this dynamic remodeling. Hypothesis: Artificial intelligence (AI)- models trained on temporal ECG feature sequences will reveal dynamic biomarkers of cardiac arrest. . Methods: Our registry consisted of 108,704 12-lead ECG recordings from 2,837 patients (age=61 ± 14 years, 34% female, 46% white) who experienced cardiac arrest within 1 year. Routine ECG features (e.g. P-duration, QRS-duration, QT-interval) were indexed by “time to arrest”. Training, test and validation cohorts were created using patient-stratified splitting, and training data was augmented to make the transformer model robust against noisy timestamps and data sparsity. We calculated the importance of features dynamically over time using Harrell’s Concordance Index (C-Index) and global SHAP values at multiple prediction horizons (30-360 days). (Fig 1) Results: In the training cohort, SCA prediction performance peaked at 60 days before arrest (C-index = 0.89) and tapered monotonically thereafter. In the hold-out test cohort, SCA prediction peaked at 120 days (C-index = 0.78) and fell only slightly to 0.76 at 365 days (Fig 2). Features most important for long-term prediction were the RR-interval (importance ~50%), with an emergence of QT-interval, P-duration and PR-interval as additional co-predictors at timepoints closer to the SCA event (cumulative importance ~ 32%) (Fig 3). Conclusion: ECG-features’ vary in their importance to predicting SCA over a one-year period. Using time-aware AI-transformer models, RR interval was the strongest predictor from 1 year preceding the event, with other features emerging as the event draws closer. This study shows the importance of using repeated 12-lead ECGs in at-risk patients, and combining them into dynamic-predictive indices.