Yale AI Uses Smartwatch ECGs to Detect Structural Heart Disease in Prospective Study

Overview

• The model was trained on 266,054 12-lead ECGs matched to echocardiograms from 110,006 patients at Yale New Haven Hospital between 2015 and 2023.
• External validation included 44,591 adults from four community hospitals and 3,014 participants from the ELSA-Brasil population study.
• In smartwatch testing, overall performance was about 88% for detecting multiple structural diseases versus 92% when using single-lead signals captured on hospital equipment.
• Only about 5% of the 600 participants had echocardiogram-confirmed disease—15 with low ejection fraction, five with severe left-sided valvular disease, and one with severe left ventricular hypertrophy—limiting precision of predictive values.
• Researchers added simulated noise during training to bolster robustness for real-world signals, and the findings are being presented at AHA Scientific Sessions 2025 with plans for larger, more diverse evaluations and workflow studies.