The ECG2Stroke deep learning model predicted 10-year ischemic stroke risk with moderate discrimination (AUC 0.745-0.791), performing similarly to the Framingham Stroke Risk Profile.
Cohort (n=205,033)
Yes
Does an AI-enabled 12-lead ECG deep learning model accurately predict 10-year incident ischemic stroke risk compared to the Framingham Stroke Risk Profile?
An AI-enabled 12-lead ECG model can predict 10-year ischemic stroke risk with accuracy comparable to established clinical risk scores, potentially aiding in stroke prevention prioritization.
Effect estimate: HR 2.17 (95% CI 1.64-2.87)
Absolute Event Rate: 0.791% vs 0.779%
BACKGROUND: Scalable risk stratification for ischemic stroke remains an unmet need. OBJECTIVES: In this study, the authors sought to assess whether deep learning of 12-lead electrocardiograms (ECGs) can estimate longitudinal ischemic stroke risk and quantify the extent to which risk signals reflect plausible mechanisms (eg, atrial cardiopathy). METHODS: We trained a convolutional neural network to estimate the 10-year risk of incident ischemic stroke with the use of 12-lead ECG among patients receiving longitudinal care at Massachusetts General Hospital (MGH). Neural network-derived stroke probabilities, age, and sex were integrated into a Cox proportional hazards model ("ECG2Stroke"). Within an MGH test set ("MGH Test"), as well as independent samples from Brigham and Women's Hospital (BWH) and Beth Israel Deaconess Medical Center (BIDMC), we assessed model discrimination (area under the curve AUC) and calibration (integrated calibration index ICI). ECG2Stroke was compared with the revised Framingham Stroke Risk Profile (FSRP). Saliency mapping, associations with clinical factors and structured ECG features, and performance across stroke subtypes were assessed. RESULTS: ECG2Stroke was developed in 101,496 individuals from MGH (age 57 ± 16 years, 48% women), and evaluated in MGH Test (n = 4,771; age 57 ± 16 years, 49% women), BWH (n = 68,884; age 57 ± 16 years, 55% women), and BIDMC (n = 29,882; age 56 ± 17 years, 54% women). At 10 years, there were 346 stroke events in MGH Test, 3,209 in BWH, and 1,236 in BIDMC. ECG2Stroke demonstrated moderate discrimination of incident stroke (10-year AUCs: MGH Test, 0.795; BWH, 0.774; BIDMC, 0.772) and low calibration error (ICIs: MGH Test, 0.030; BWH, 0.005; BIDMC, 0.026). In patients with available data, 10-year AUC for ECG2Stroke was similar to FSRP (MGH/BWH Test: ECG2Stroke, 0.791; FSRP, 0.779; BIDMC: ECG2Stroke, 0.745; FSRP, 0.728). Stratification persisted across subgroups, including patients with and without atrial fibrillation. Saliency maps highlighted the ECG P-wave, and risk estimates correlated with structured P-wave indices. ECG2Stroke was strongly associated with cardioembolic stroke (cause-specific HR: 2.17 per 1-SD of logit-transformed probability; 95% CI: 1.64-2.87) but not noncardioembolic stroke. CONCLUSIONS: ECG-based artificial intelligence (AI) can predict 10-year ischemic stroke with performance similar to a validated clinical score, possibly by encoding markers of abnormal atrial substrate linked to cardioembolism. AI-enabled ECG analysis may enable efficient prioritization for stroke prevention.
Mahajan et al. (Fri,) conducted a cohort in Ischemic stroke risk (n=205,033). ECG2Stroke (deep learning model of 12-lead ECG) vs. Revised Framingham Stroke Risk Profile (FSRP) was evaluated on 10-year risk of incident ischemic stroke (model discrimination AUC) and cardioembolic stroke (HR 2.17, 95% CI 1.64-2.87). The ECG2Stroke deep learning model predicted 10-year ischemic stroke risk with moderate discrimination (AUC 0.745-0.791), performing similarly to the Framingham Stroke Risk Profile.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: