Cephalalgia · Nov 17, 2026
Vascular risk prediction in migraine: A multimodality risk score based on artificial intelligence-electrocardiogram output, echocardiogram, and detailed migraine characteristics
By Ihara K, Chiang CC, Yang PH et al.
Aim: Patients with migraine, especially those with aura, have a higher risk of stroke and other vascular events. However, the underlying mechanisms are complex, multifactorial, and may require integration of multimodal data for comprehensive characterization. We have conducted prior analyses on how artificial intelligence (AI)-enabled electrocardiogram (ECG), specifically the atrial fibrillation prediction model and age estimation model output, echocardiogram parameters, and detailed migraine characteristics separately predict vascular events. Herein, we aim to investigate the relative importance and combined effects of all three modalities for vascular risk prediction. Methods: In this retrospective cohort study, we developed a multimodality prediction risk score to evaluate the vascular risks in patients with migraine based on the coefficients of variables extracted from our prior studies, including comprehensive AI-ECG atrial fibrillation prediction model output and computed delta age (ECG estimated age minus chronological age), echocardiogram parameters, and migraine characteristics. We then included patients with at least one 12-lead ECG, transthoracic echocardiogram, and a visit with a headache specialist at any Mayo Clinic site within three years to evaluate the risk score. The outcome is a composite endpoint of acute ischemic stroke, acute myocardial infarction, venous thromboembolism, and all-cause mortality. We selected 20 risk parameters, including demographics, comorbidities, AI-ECG outputs, echocardiogram features, and migraine characteristics found to be predictive of vascular events. Each variable was assigned a different risk score according to the hazard ratio (HR) and beta coefficient from prior Cox regression analysis for each modality. We used Contal and Q’Quigley's method to find the optimal cut-off, and an additional multivariate Cox regression was conducted to assess the relative risk of each parameter category in the final cohort. We investigated the added value of each modality by net reclassification improvement rate. Results: A total of 29,928 patients, 12,778 patients, and 6819 patients were included to investigate how AI-ECG outputs, echocardiogram parameters, and migraine characteristics, respectively, predicted adverse vascular outcomes, and the results were used to develop a multimodal risk score. A total of 297 patients with migraine included in previous analyses, of whom 247 (83.2%) were female and 109 (36.7%) had aura, were used for proof-of-concept evaluation of the developed risk score. The mean (standard deviation, SD) age was 39.0 (13.3) and follow-up period was 68.9 (60.9) months. Thirty-five patients (11.8%) developed the composite endpoint. Those with a risk score ≥15 had a significantly shorter vascular event-free survival (HR [95% Confidence Interval (CI)], 3.83 [1.89–7.76]; p < 0.001) compared to those with a risk score <15 in our cohort of 297 patients. Furthermore, comorbidity (HR, 1.17 [1.06–1.30]; p = 0.003) and migraine characteristics (HR, 1.33 [1.06–1.67]; p = 0.014) were significantly predictive of the composite endpoint, whereas other categories were not. Net reclassification improvement was 17.7% (95% CI, 15.9−19.9%; p < 0.001) for AI-ECG model output, 16.9% (95% CI, 14.1−19.5%; p < 0.001) for echocardiogram, and 22.1% (95% CI, 16.7−26.9%; p < 0.001) for migraine characteristics, suggesting significant contribution from each modality. Conclusion: This pilot evaluation study supports the use of combined AI-ECG output, echocardiogram, and migraine characteristics for vascular risk prediction. Although future external validation and real-world clinical assessment is necessary, our preliminary results suggest significant added value of each modality and indicate the importance of obtaining detailed migraine features for vascular risk stratification.