Transformer-based ECG foundation and deep learning models for assessment of left ventricular systolic dysfunction

Abstract Background Early identification or rule-out of left ventricular systolic dysfunction (LVSD) is critical for optimizing clinical care pathways. AI-enabled analysis of 12-lead resting ECGs offers a low-cost and scalable solution, but direct comparisons between foundation models and conventional deep learning architectures remain limited. Purpose To evaluate the performance of a novel time-series foundation model (TSFM) for detecting LVSD from ECGs and compare it with a standard ResNet-18 deep learning (DL) model. Methods Both models were trained and tested using 31,832 Heart Center Leipzig ECG–echocardiogram pairs, split into training (n = 20,372), validation (n = 5,093), and test (n = 6,367) cohorts. TSFM was additionally pre-trained on 33K unlabeled ECGs using masked pre-training. LVSD was defined as LVEF ≤ 40%, with a prevalence of 15% across all sets. Model performance was evaluated using AUROC, AUPRC, and standard classification metrics. Results On the test dataset, TSFM achieved a higher AUROC (0.924 [95% CI: 0.918–0.928]) than ResNet-18 (0.912 [95% CI: 0.904–0.916], p < 0.001), and a slightly higher but not statistically significant AUPRC (0.695 [0.663–0.723] vs. 0.646 [0.612–0.683]). TSFM also demonstrated improved sensitivity (0.89 vs. 0.87) and negative predictive value (0.98 vs. 0.97), with identical specificity (0.81) and positive predictive value (0.44). Conclusions The transformer-based ECG model outperforms a conventional DL model in detecting LVSD. Its high negative predictive value supports further evaluation in triage, emergency, and preoperative settings where rapid rule-out of LVSD can inform downstream clinical decision-making.