Explainable electrocardiogram-based deep learning predicts ventricular arrhythmias after cardiac resynchronisation therapy

Introduction Cardiac resynchronisation therapy (CRT) improves cardiac function and clinical outcomes in dyssynchronous heart failure. Despite reverse remodelling, ventricular arrhythmias (VAs) may still occur and arrhythmic risk stratification remains challenging. Deep neural networks can identify complex ECG patterns associated with diverse outcomes, but their clinical adoption is limited by their “black-box” nature. Explainable deep learning approaches such as FactorECG may facilitate clinical implementation. Although FactorECG has demonstrated predictive value for CRT response and adverse outcomes, its ability to stratify arrhythmic risk after CRT remains unclear. Purpose To evaluate the discriminative performance of an explainable ECG-based deep learning algorithm for VAs after CRT and to identify associated ECG features. Methods From a multicentre prospective cohort of 240 CRT recipients (2012–2013), patients meeting contemporary guideline-based criteria with long-term follow-up and an analysable baseline ECG were included. An explainable deep learning algorithm based on variational auto-encoders, pretrained on 1.1 million ECGs from 251,473 patients, was used to compress median-beat ECGs into 21 interpretable factors (FactorECG). Cox proportional hazards models assessed the association between FactorECG and VAs (ventricular tachycardia or fibrillation) after CRT. To mitigate overfitting, ridge-penalised Cox regression was applied, with internal validation using bootstrap-derived optimism-corrected C-indices. The relative contribution of individual factors was explored with cross-validated LASSO. Patients were stratified by median predicted risk and visualised using Kaplan–Meier analysis. Results The final study population comprised 198 patients (mean age 66 ± 10 years; 64% male; 45% ischaemic cardiomyopathy). During a median follow-up of 9.2 years (IQR 4.8–11.5), 55 patients (28%) experienced VAs. FactorECG predicted VAs with an optimism-corrected C-index of 0.71 (95% CI 0.65–0.77). The strongest associations were observed for latent features representing prolonged PR interval and P-wave duration, increased P-wave amplitude, reduced precordial QRS amplitude, lateral or inferolateral ST-segment elevation, and shorter QT duration. Compared with low-risk patients, high-risk patients more frequently had ischaemic cardiomyopathy (59% vs 32%), less often had LBBB (42% vs 85%), and demonstrated lower rates of reverse remodelling at 6 months (37% vs 75%) (all p<0.001). FactorECG remained predictive among patients with reverse remodelling at 6 months (optimism-corrected C-index 0.69; 95% CI 0.59–0.78). Conclusion Based solely on a standard 12-lead ECG, FactorECG demonstrated promising discriminative performance for VAs after CRT, including among patients with reverse remodelling. This approach may refine arrhythmic risk stratification and inform consideration of defibrillator capability. External validation is warranted.