ECGxAI

ECGxAI predicts outcomes after cardiac resynchronization therapy (CRT) by applying explainable deep learning to median-beat 12-lead electrocardiogram (ECG) data to identify ECG features associated with CRT response.

Key Features:

• Deep learning-based compression: An explainable deep learning model compresses median-beat 12-lead ECGs into 21 interpretable factors termed FactorECG that summarize critical ECG features.
• Training dataset scale: The model was trained on 1.1 million ECGs from over 251,473 patients.
• Predictive performance: In clinical validation on pre-implantation ECGs from 1,306 CRT patients across three academic centers, FactorECG achieved a c-statistic of 0.69 (95% CI 0.66–0.72), outperforming guideline-based ECG criteria (c=0.57, 95% CI 0.54–0.60) and QRSAREA (c=0.61, 95% CI 0.58–0.64).
• Explainable predictors: Identified key predictors associated with poor CRT outcomes include inferolateral T-wave inversion, reduced right precordial S- and T-wave amplitude, ventricular rate, increased PR interval, and prolonged P-wave duration.
• Input data requirement: Model predictions rely solely on a standard 12-lead ECG median beat without additional clinical variables.

Scientific Applications:

• CRT outcome prediction and patient selection: Predicts individual likelihood of CRT response to inform selection of candidates for therapy.
• ECG biomarker discovery: Identifies and quantifies ECG features (FactorECG components) associated with CRT outcomes for biomarker development and mechanistic investigation.
• Comparative evaluation of ECG metrics: Provides a benchmark for comparing deep-learning–derived ECG factors against guideline-based ECG criteria and QRSAREA.

Methodology:

An explainable deep learning model was trained on 1.1 million median-beat 12-lead ECGs from >251,473 patients to compress ECGs into 21 interpretable FactorECG components, with performance validated on pre-implantation ECGs from 1,306 CRT patients across three academic centers and reported using c-statistics (95% CIs) compared to guideline criteria and QRSAREA.