Abstract Body: INTRODUCTION: Echocardiography is the reference standard for structural heart disease (SHD), but access limits population screening. Twelve-lead electrocardiography (ECG) is scalable, yet the optimal transfer strategy for ECG foundation models in echo-confirmed multi-label SHD detection is unclear.
HYPOTHESIS: Domain adaptation followed by selective fine-tuning would outperform engineered features, de novo waveform modeling, and frozen-encoder transfer.
AIMS: To develop a compute-aware ECG framework for six moderate-or-greater SHD phenotypes and define the performance-efficiency trade-off of selective backbone adaptation.
METHODS: Using EchoNext Mini-Model (100,000 ECG-echocardiogram pairs), we modeled reduced left ventricular (LV) ejection fraction, right ventricular (RV) dysfunction, increased LV wall thickness, aortic stenosis, mitral regurgitation, and tricuspid regurgitation. We compared XGBoost, ResNet1D, frozen ECG-FM, and a two-stage ECG-FM strategy with continued self-supervised pretraining on EchoNext waveforms followed by layer-wise partial fine-tuning.
RESULTS: Domain-adapted partial fine-tuning performed best. With four frozen layers, macro AUROC/AUPRC was 0.8509/0.4261 versus 0.8017/0.3289 for XGBoost, 0.8246/0.3833 for ResNet1D, and 0.8233/0.3670 for frozen ECG-FM. With eight frozen layers, fixed-threshold F1 was highest at 0.3691 while updating 41.46M parameters, exceeding full fine-tuning (0.3423) and linear probing (0.1933). Late fusion with seven covariates did not improve AUROC/AUPRC, and LoRA underperformed partial fine-tuning (0.8196/0.3523; F1=0.1963). The best adapted single backbone also outperformed ECGFounder and HuBERT-ECG, and unified multi-label training outperformed six task-specific binary models (AUROC=0.8501 vs 0.8459; AUPRC=0.4173 vs 0.4111; F1=0.3691 vs 0.2769). Downsampling all-negative training records increased F1 to 0.4103 at the cost of lower AUROC/AUPRC, providing a simple operating-point knob.
CONCLUSIONS: Domain-adapted partial fine-tuning improved multi-label ECG screening for echo-confirmed SHD and offered a favorable performance-efficiency trade-off, supporting selective adaptation as a practical default for compute-aware triage.