External validation of an explainable electrocardiogram-only deep learning algorithm for the prediction of response...

Cardiac resynchronization therapy (CRT) can improve clinical outcomes in patients with dyssynchronous heart failure, but many patients selected according to the current guidelines do not respond.

This study aimed to externally validate an explainable deep learning algorithm (the FactorECG algorithm) for predicting response after biventricular pacing.

We previously trained a deep learning algorithm on >1 million electrocardiogram (ECG) median beats to learn the underlying generative factors of the ECG and applied it to 1306 patients with CRT from the Netherlands. Using the extracted 21 explainable factors, a model predicting the risk of volumetric nonresponse and poor clinical outcomes was developed. In the present analysis, this model was externally validated in a cohort of 161 patients with CRT from the University of Virginia for volumetric nonresponse only. Subsequently, the added value of clinical and cardiac magnetic resonance imaging-derived predictors was investigated.

The original model significantly outperformed American Heart Association criteria for left bundle branch block for the prediction of nonresponse {C-statistic 0.67 (95% confidence interval [CI] 0.59-0.76) vs 0.51 (95% CI 0.41-0.60), respectively}. A refitted FactorECG-based model performed similarly to a model also integrating indices of mechanical dyssynchrony (C-statistic 0.74 [95% CI 0.66-0.82] vs 0.70 [95% CI 0.62-0.77], respectively). A combination of both models improved response prediction (C-statistic 0.79 [95% CI 0.71-0.85]).

In this external validation study, an explainable ECG-only algorithm for the prediction of nonresponse after CRT device implantation generalized well to a lower-risk population from a different hospital. Adding indices of mechanical dyssynchrony and right ventricular function might be of additional value when evaluating volumetric response.