Deep Learning-Based Prediction Model for Cardiac Resynchronization Therapy Responders Using Electrocardiogram Data

Comprehensive Summary

This single-center retrospective study evaluated whether preimplantation 12‑lead ECGs can predict response to cardiac resynchronization therapy (CRT) using deep learning. Among 346 CRT implants (2007–2024), 285 patients with 6‑month follow‑up and available ECGs were analyzed; responders were defined as having > 15% reduction in LV end‑systolic volume. Three models were trained: a ResNet‑18 image model, a self‑supervised learning (SSL) + ResNet‑18 image model, and a LightGBM model using summary time‑series ECG features. Across 10 random splits, the LightGBM model had the highest mean accuracy (64.7%), but the SSL + ResNet‑18 model provided the best balance of accuracy (62.1% ± 5.4), PPV (78.5% ± 5.5), and stability. In the best run, AUCs were 0.766 for SSL‑ResNet‑18, 0.728 for ResNet‑18, and 0.701 for LightGBM. NPV was around 50% for all models

Outcomes and Implications

If validated externally, an ECG‑based model could help identify likely CRT responders before implantation, potentially supporting more confident device decisions and closer follow‑up of predicted responders. However, the modest accuracy and low NPV mean these tools should not be used to deny CRT to guideline‑eligible patients; instead, they might flag patients for early optimization or alternative pacing strategies when predicted nonresponse is high. Practical deployment is possible, but generalizability is uncertain given the single‑center sample (n=285) and different CRT techniques.