Abstract
BACKGROUND: Cardiac resynchronization therapy (CRT) is an established treatment for advanced heart failure, but approximately 30% of patients fail to respond. This study aimed to develop and evaluate deep learning models using preimplantation electrocardiogram (ECG) data to predict CRT response. METHODS: We conducted a retrospective analysis of 285 patients who underwent CRT implantations and completed a 6-month follow-up. Responders were defined as those exhibiting ≥ 15% left ventricular end-systolic volume reduction. Three models were developed: ResNet-18 model trained on ECG images, self-supervised learning (SSL) enhanced ResNet-18 model, and LightGBM model trained on time-series ECG data. Model performance was evaluated using accuracy, positive predictive value (PPV), and negative predictive value (NPV), averaged across 10 random seeds. Model interpretability using Gradient-weighted Class Activation Mapping (Grad-CAM) was performed on 36 responder cases. RESULTS: The SSL + ResNet-18 model demonstrated the most stable performance (accuracy 78.5% ± 5.5%) and PPV of 81.3%. The ResNet-18 model achieved the highest PPV of 84.2% but had lower accuracy (74.1%) and larger variability. The LightGBM model exhibited the highest accuracy (79.4%) but the lowest PPV at 72.8%. Grad-CAM showed that precordial leads were highlighted in 13 cases (36.1%), limb leads in 16 (44.4%), and both regions in 7 (19.4%), indicating heterogeneity in the model's focus and potential diversity in the electrical features contributing to CRT response prediction. CONCLUSION: AI models using preimplantation ECG data, particularly those based on image inputs, can effectively predict CRT responders. This approach may enhance patient selection and support personalized therapy strategies in CRT management.