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

Background ECG foundation models (FMs) are emerging for screening or rule-out of left ventricular systolic dysfunction (LVSD). However, previous studies rely on multiple ECGs per patient, lack head-to-head comparisons with conventional deep learning (DL) models, and rarely report on short- and long-term reproducibility. Purpose To compare a de novo time-series FM (TSFM) with a DL model (ResNet-18) for LVSD detection (LVEF≤40%), evaluate robustness across EF thresholds and subgroups, quantify short- and long-term reproducibility, assess external generalisability and prognostic utility. Methods TSFM was pretrained on unlabeled ECG data (N=189,886 from our hospital ECG archive and MIMIC-IV) using masked pre-training and fine-tuned for LVSD classification based on ECG-echocardiogram pairs. ResNet-18 was trained under the same endpoint definition. Performance was evaluated on a held-out test set (N=10,582) and in external validation (EchoNext, N=32,660). Threshold robustness (LVEF 20-50%) was assessed without retraining by relabeling. Subgroup analyses were performed in MIMIC-IV. Across pretraining, fine-tuning, and primary evaluation, one ECG per patient was used. Test–retest reproducibility was assessed in ECG pairs with unchanged echocardiography-defined LVSD status (short-term 1–14 days, N=16,503 ECG pairs; long-term ≥365 days, N=854). Prognostic utility was assessed in patients with baseline preserved LVEF (≥50%) over 5-year follow-up. Results LVSD prevalence was 16-20% across all sets. In the test set, TSFM showed higher AUROC than ResNet-18 (0.896 [0.887-0.902] vs 0.885 [0.876-0.892]), DeLong p<0.001. AUPRC was also higher (0.695 [0.663-0.723] vs 0.646 [0.612-0.683]), permutation p=0.051. Sensitivity was 0.82 for both, while TSFM showed higher specificity (0.82 vs 0.79) and PPV (0.53 vs 0.49) with NPV 0.95 for both. TSFM maintained higher AUROC across LVEF thresholds. In MIMIC-IV subgroups, TSFM showed higher AUROC across age, sex, heart rate, rhythm, and QRS duration. Short-term agreement was 83.3% for TSFM vs 87.2% for ResNet-18 (κ 0.655 vs 0.671; Δκ p=0.0796) and long-term agreement of 76.0% vs 82.1% (κ 0.520 vs 0.612; Δκ p=0.013). In EchoNext, TSFM achieved AUROC 0.875 (0.870-0.880) vs ResNet-18 0.850 (0.844-0.857), DeLong p<0.001; AUPRC 0.640 (0.627-0.653) vs 0.592 (0.578-0.605), permutation p<0.001. In baseline preserved-EF patients, model-flagged LVSD positivity predicted subsequent LVSD over 5 years (TSFM HR 2.96 [2.44-3.58]; ResNet-18 HR 3.44 [2.79-4.25]); interaction HR 0.86 (0.65-1.14), p=0.30. Conclusions In head-to-head testing, TSFM consistently improved LVSD detection across EF thresholds, clinical subgroups and external validation, with good overall reproducibility. High NPV at the observed prevalence supports evaluation for rapid LVSD rule-out in triage, emergency, and preoperative workflows. Prognostic stratification for future LVSD seems feasible.Performance metricsLVSD prediction