Machine-learned physiological signatures from the ECG predict sudden death in ischemic cardiomyopathy

Funding Acknowledgements Type of funding sources: Public grant(s) – National budget only. Main funding source(s): National Institute of Health (NIH) Background Low left ventricular ejection fraction (LVEF) is an imperfect predictor of sudden cardiac death (SCD) in patients with ischemic cardiomyopathy. Novel features from the ECG might provide a readily available tool to better predict risk. Purpose We hypothesized that machine learning (ML) of the ECG can be used to predict SCD, and the ML-learned ECG features could be referenced to interpretable intracardiac signals (monophasic action potentials: MAP) to provide mechanistic insights. Methods We studied 5603 ECG Lead V1 beats in 41 patients (64±10 Y) with coronary disease and LVEF≤40% in steady-state pacing. Patients were randomly allocated to independent training and test cohorts in a 70:30 ratio, repeated K=10-fold. Support vector machines were trained to predict mortality at 3Y from the top 20 features derived from these beats. Patient-level predictions were made by computing an ECG score that indicates the proportion of test set beats in that patient computed by the beat-level model to predict death. Explainability analysis was performed using the arithmetic mean of MAP and ECG beats that predicted SCD versus those that predicted survival. Results Fig 1A. shows ECG lead V1 and MAP in a 79 Y man with LVEF 29%. Fig 1B shows the dataflow in the study. Predictive accuracies of ML models were 78 and 70% and optimal with 20 features for both ECG and MAP models respectively (Fig. 1C). Beat-level predictions in the validation (n=1678 Lead I beats) cohorts yielded c-statistics of 0.78 with the ECG (95% CI, 0.62–0.91) and 0.75 with MAPs (95% CI, 0.75-0.76) (data not shown). In multivariable patient-level models, c-statistic was 0.87 with ECGs (95% CI, 0.76-0.98) (Fig 1D) and 0.82 with MAPs. On explainability analysis, ECG beats that predicted SCD (Fig 2; red) had lower amplitude and more notched T-waves in lead V1 than beats that predicted no SCD (Fig 2; blue). MAP that predicted SCD had higher repolarization current at the same time points. Both QT duration (ECG) and action potential duration (MAP) did not differ (Fig 2). Conclusions Machine learning of the ECG reveals novel predictors of SCD risk in patients with ischemic cardiomyopathy analogous to those identified in intracardiac signals. This approach can be used as a point-of-care ECG risk tool to improve risk stratification and allocation for ICD therapy beyond LVEF alone and may shed insights into the pathophysiology of ventricular arrhythmias.