Efficient approaches for long-term ischemic stroke risk stratification remain clinically important, particularly in patients without established cerebrovascular disease. A study published in the Journal of the American College of Cardiology evaluated whether artificial intelligence (AI)-enabled analysis of routine 12-lead electrocardiograms (ECGs) can estimate 10-year ischemic stroke risk.

The analysis included patients receiving longitudinal care at Massachusetts General Hospital (MGH), where a convolutional neural network was trained using ECG data from 101,496 individuals. Neural network–derived stroke probabilities were integrated with age and sex into a Cox proportional hazards model termed ECG2Stroke and externally validated in cohorts from Brigham and Women’s Hospital (BWH) and Beth Israel Deaconess Medical Center (BIDMC).

Findings

• The derivation cohort included 101,496 individuals from MGH, while validation cohorts included MGH Test (n=4,771), BWH (n=68,884), and BIDMC (n=29,882).
• At 10 years, ischemic stroke events occurred in 346 participants in MGH Test, 3,209 in BWH, and 1,236 in BIDMC.
• ECG2Stroke demonstrated moderate discrimination for incident ischemic stroke, with 10-year AUCs of 0.795 in MGH Test, 0.774 in BWH, and 0.772 in BIDMC.
• Calibration error remained low across cohorts, with integrated calibration indices of 0.030 in MGH Test, 0.005 in BWH, and 0.026 in BIDMC.
• In participants with available comparator data, ECG2Stroke performed similarly to the revised Framingham Stroke Risk Profile (MGH/BWH Test: 0.791 vs 0.779; BIDMC: 0.745 vs 0.728).
• The model demonstrated stronger associations with cardioembolic stroke (HR per 1-SD increase in logit-transformed probability: 2.17; 95% CI 1.64–2.87) than with noncardioembolic stroke, while maintaining risk stratification in patients with and without atrial fibrillation.

The findings suggest that AI-enabled ECG analysis may help estimate long-term ischemic stroke risk using routinely acquired cardiovascular data.