Deep Learning Model Using Transfer Learning for Detecting Left Ventricular Systolic Dysfunction: Retrospective Algorithm Development and Validation Study

Background: Artificial intelligence-augmented electrocardiogram (AI-ECG) models for detecting left ventricular systolic dysfunction (LVSD) often exhibit degraded performance in patients with comorbidities. Objective: This study aimed to introduce and validate a recalibration method using longitudinal patient data to enhance prediction accuracy and simulate its clinical utility for ongoing monitoring. Methods: We conducted a multicenter, retrospective cohort study using data from 2 hospitals in Korea. A dataset of paired transthoracic echocardiograms (TTEs) and electrocardiograms (ECGs) matched within a 2-week interval was constructed, separating pairs into baseline (first for each patient) and follow-up assessments. In addition to conventional supervised learning, we developed a patient-wise recalibration strategy that incorporated historical left ventricular ejection fraction measurements and prior AI-ECG outputs to adjust for future predictions, thus empirically mitigating confounding effects. Pretraining was also implemented to enhance the model's performance. Results: The recalibrated 12-lead DeepECG LVSD model achieved an area under the receiver operating curve of 0.956 (95% CI 0.946-0.965) for internal validation and 0.940 (95% CI 0.936-0.945) for external validation of follow-up TTE-ECG pairs. The uncalibrated 12-lead DeepECG LVSD model also showed modest performance, with an area under the receiver operating curve of 0.953 (95% CI 0.941-0.965) in the internal validation and 0.947 (95% CI 0.943-0.951) in the external validation when tested on baseline TTE-ECG pairs. Recalibration yielded statistically significant improvements in the 12-lead DeepECG LVSD models (P