AI-powered wearable ECG system detects myocardial ischemia minutes earlier

Myocardial ischemia, the primary driver of heart attacks, remains the leading cause of death and disability worldwide. Delays in diagnosis directly correlate with increased myocardial necrosis, higher complication rates, and elevated mortality. While traditional 12-lead ECG is the clinical gold standard for ischemia detection, its episodic nature fails to capture transient, unpredictable ischemic episodes during continuous ambulatory monitoring. Though wearable ECG devices have excelled at detecting arrhythmias like atrial fibrillation (with over 95% sensitivity), their utility for ischemia detection has long been limited by the subtle, multiscale temporal changes in ECG waveforms that mark ischemic progression-from minute alterations in the ST-segment and T-wave to beat-to-beat variability shifts that evolve over minutes to hours.

To address these longstanding clinical gaps, the research team developed a hierarchical temporal fusion transformer architecture that models ischemic dynamics across three physiologically critical timescales simultaneously: intra-beat morphological feature extraction to capture the earliest ischemic markers, inter-beat variability modeling to track cardiac stress progression, and long-term trend analysis via dilated temporal convolutional networks. The framework uses dual-task learning to jointly predict impending ischemia and stratify post-reperfusion injury risk, boosting performance through shared pathophysiological representations. The system pairs with an FDA-cleared, chest-worn single-lead ECG patch with 14-day continuous monitoring capability and over 92% signal quality acceptance during daily activities.

Validated across four large-scale datasets encompassing 108,778 total patients (including 17,173 ischemia-positive cases), the framework achieved an overall area under the receiver operating characteristic curve (AUROC) of 0.947 for ischemia detection, a 4.8% to 9.5% relative improvement over state-of-the-art baseline models. It reached 84.1% to 87.3% sensitivity at 90% specificity across all cohorts, with a concordance index (C-index) of 0.923 for post-reperfusion risk stratification. Critically, the system maintained an 88.7% positive predictive value (PPV) at 15 minutes and 84.1% at 20 minutes, ensuring reliable, actionable alerts with minimal false positives to avoid clinician alert fatigue. It delivered consistent performance across all age, sex, and comorbidity subgroups, with no evidence of demographic bias. For real-world deployment, the full model runs inference on 10-second ECG segments in just 47.3 ms, with a lightweight pruned variant reducing inference time to 28.6 ms while retaining an AUROC above 0.93, making it compatible with standard clinical hardware.

The 18.4-minute early warning window addresses the core clinical challenge of "time is muscle" in acute coronary syndrome management, enabling bedside assessments, emergency protocol initiation, and catheterization lab mobilization before irreversible myocardial damage occurs. The system's attention patterns align closely with cardiologist-identified ischemic markers (Spearman correlation 0.78–0.84), ensuring strong clinical interpretability. Limitations include the predominantly Chinese hospital-based study cohorts, requiring further validation across diverse ethnic, socioeconomic, and healthcare settings, alongside prospective clinical trials to confirm real-world patient outcomes. Future research will expand the framework to predict additional cardiovascular events, integrate electronic health records for personalized risk assessment, and develop federated learning approaches to protect patient privacy while refining model performance.

Authors of the paper include Songtao An, Jiamin Yuan, Yang Pan, Miaoqing Ye, Zhenghan Chen, Minying Li, Panyue Yan, Jiali Yao, Yujie Guan, Yan Lin, Wenjuan Wang, Haliminai Dilimulati, Yuanyin Teng, Keyu Dai, Yuqi Bai, Junbo Ge, and Dong Deng.

This research was partially supported by the Youth Programme of National Natural Science Foundation of China (No. 82304983), the National Natural Science Grant of China (No. 81970312), the Soochow University Horizontal Project (Code Numbers: H230269 and H240140), and the Multi-center Clinical Research Project for Major Diseases in Suzhou (Grant Number: DZXYJ202302). This research is supported by the Undergraduate Innovation Laboratory of School of Pharmaceutical Science, Guangzhou University of Chinese Medicine.