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

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Cardiac resynchronization therapy (CRT) improves outcomes in heart failure patients with dyssynchronous contractions.
• However, a significant portion of patients selected by current guidelines do not respond to CRT.
• Predicting CRT response remains a clinical challenge.

Purpose of the Study:

• To externally validate the FactorECG pipeline, an explainable deep learning model, for predicting response to biventricular pacing (BIVP).
• To assess the performance of the FactorECG model in a new patient cohort.
• To investigate the added value of clinical and cardiac imaging predictors.

Main Methods:

• A deep learning algorithm trained on over 1 million ECGs was used to extract 21 explainable factors.
• A predictive model for volumetric non-response and poor outcome was developed using these factors.
• The model was externally validated in 161 CRT patients from the University of Virginia.

Main Results:

• The FactorECG model significantly outperformed AHA criteria for Left Bundle Branch Block (LBBB) in predicting non-response (c-statistic 0.67 vs. 0.51).
• A refitted FactorECG model integrating mechanical dyssynchrony indices showed comparable performance (c-statistic 0.74 vs. 0.70).
• Combining both models improved response prediction (c-statistic 0.79).

Conclusions:

• The explainable ECG-only FactorECG algorithm demonstrated good generalization for predicting CRT non-response in an external, lower-risk population.
• The model's performance was robust across different healthcare settings.
• Integrating mechanical dyssynchrony and right ventricular function data may further enhance volumetric response prediction.