Bridging the gap from clinical to home ECG: quantifying and overcoming accuracy loss in AI-enabled single-lead ECG...

AI-enabled electrocardiogram (ECG) models trained on 12-lead recordings may underperform on home single-lead devices due to device and context domain shift. We trained Lead-I models using 676,192 ECGs from 246,874 patients and externally validated them in 219,231 ECGs from 65,338 patients. We evaluated two methodological control indicators (gender and age) and seven clinical phenotypes [90-day mortality, left ventricular ejection fraction (EF), pulmonary artery systolic pressure (PASP), left atrial diameter, N-terminal pro-B-type natriuretic peptide (NT-proBNP), hemoglobin, and estimated glomerular filtration rate]. Cross-hospital validation showed minimal overall degradation (AUC decline <0.03). In contrast, direct deployment to consumer ECGs from Apple Watch (n = 9835) and QOCA ECG102D (n = 31,517) produced substantial accuracy losses. After fine-tuning, AUCs across eight classification tasks for the Apple Watch and QOCA ECG102D were significantly improved, with high performance for low EF (0.85/0.86) and more modest performance for elevated PASP (0.73/0.68) and anemia (0.73/0.69). Performance was higher in sinus rhythm than atrial fibrillation. Learning curves indicated that approximately 500-1000 labeled ECGs are required for reliable fine-tuning; smaller sets can be harmful. We release an evaluation-only HOME dataset to facilitate reproducible benchmarking of cross-device generalization. Models should be interpreted as screening or triage tools, not standalone diagnostic tests.