Abstract 4362170: Artificial Intelligence-Enabled Electrocardiogram for Detecting Systolic Dysfunction in Left Bundle Branch Block: Impact of Training Strategies and Transfer Learning

Abstract

Background: Left bundle branch block (LBBB) is a common ECG abnormality linked to left ventricular systolic dysfunction (LVSD), yet its altered conduction pattern complicates ECG-based diagnosis. While artificial-intelligence enabled ECG (AI-ECG) show promise for LVSD screening, it is unclear whether subgroup-specific training or transfer learning improves performance. Hypothesis: We hypothesize that diagnostic performance for detecting LVSD in LBBB patients varies depending on the model development strategy, including training cohort selection and transfer learning approaches. Methods: We developed four AI-ECG models using 364,845 ECGs: (1) a general model, (2) a model trained on automatically extracted LBBB cases, (3) a model trained on expert-validated LBBB ECGs, and (4) a transfer learning model fine-tuned on expert-curated LBBB data. All models were externally validated on 1,334 ECGs from LBBB patients at a separate hospital. LVSD was defined as ejection fraction ≤40%. Performance was assessed using AUROC, sensitivity, specificity, and predictive values. Results: The transfer learning model achieved the highest AUROC (0.903; 95% CI: 0.887–0.918), followed by the general model (0.899; 95% CI: 0.883–0.915). Models trained solely on LBBB datasets performed worse (AUROC 0.879 and 0.841). The general model had the highest sensitivity (96.6%), while the transfer learning model had greater specificity (90.3%) and PPV (89.4%). ( Figure 1 ) Kaplan–Meier analysis showed that AI-ECG false positives had significantly higher risk of future LVSD than true negatives (38% vs. 8.5%, p<0.0001). ( Figure 2 ) We further explored ECG characteristics of LVSD in LBBB patients to evaluate model explainability. ( Figure 3 ) Conclusions: A general AI-ECG model effectively detects LVSD in LBBB patients, and transfer learning offers modest improvement. Model development strategies, especially training cohort composition, influence diagnostic performance in high-risk subgroups. AI-ECG holds promise for screening and prognostication in LBBB populations, warranting further validation.