AI flags 3-month MACE risk in smartwatch ECGs after decompensated HFrEF: a proof-of-concept study

Abstract Background/Introduction Patients hospitalised for acute decompensated heart failure with reduced ejection fraction (HFrEF) remain at high risk of early readmission or cardiovascular death. Single-lead electrocardiography (ECG) from consumer smartwatches offers a low-burden tool for post-discharge monitoring, but its prognostic potential is largely unexplored. Building on an AI-ECG model adapted from 12-lead to single-lead recordings and validated in clinical datasets 1, we investigated whether smartwatch ECGs recorded at discharge contain information predictive of 3-month major adverse cardiovascular events (MACE). Purpose To test the feasibility of using single-lead smartwatch ECGs at hospital discharge to identify patients with elevated 3-month MACE risk following recompensation for HFrEF. Methods In this prospective, observational study, 32 adults hospitalised with acute decompensated HFrEF were enrolled between April 2024 and July 2025. Each participant recorded 30-s single-lead smartwatch ECGs on the day of discharge; follow-up for MACE was 3 months. Mean age was 70.9 ± 15.1 years; 84.4% were male (27/32). A single-beat AI-ECG model that is trained to classify four arrhythmias (left bundle branch block (LBBB), first-degree atrioventricular block (1dAVb), atrial fibrillation (AF), right bundle branch block (RBBB)) was applied to each heartbeat. Beat-level probabilities were aggregated into patient-level arrhythmia burdens. Logistic regression models assessed associations between arrhythmia burdens and 3-month MACE (heart-failure rehospitalisation or cardiovascular death), adjusting for age, sex, NT-proBNP and left-ventricular ejection fraction (LVEF). Discrimination was evaluated by odds ratios (per 1-SD change) with bootstrap 95 % confidence intervals. Results Fifteen participants (47 %) experienced MACE (11 rehospitalisations, 4 deaths) within three months. Higher AI-derived burdens of RBBB (OR = 10.97 1.29–93.60), LBBB (OR = 4.46 0.96–20.71), 1dAVb (OR = 5.17 1.13–23.59) and AF (OR = 5.17 1.13–23.59) were associated with increased MACE risk after adjustment for age, sex, NT-proBNP, and LVEF at discharge. Associations were directionally consistent across arrhythmia types and remained similar when excluding low-quality ECGs, supporting the robustness of the signal despite limited sample size. Conclusion(s) Single-lead smartwatch ECGs recorded at discharge capture arrhythmia patterns associated with 3-month adverse outcomes in HFrEF. While limited by cohort size, these findings provide proof-of-concept that AI-derived arrhythmia features from wearables may contribute to post-discharge risk flagging. Larger studies should validate these observations and explore integration with clinical and biomarker data for individualised follow-up strategies.For image description, please refer to the figure legend and surrounding text. For image description, please refer to the figure legend and surrounding text.