Electrocardiogram-Based Artificial Intelligence to Discriminate Cardioembolic Stroke and Stratify Risk of Atrial Fibrillation After Stroke

Atrial fibrillation (AF) is the most common cause of cardioembolic stroke. Recognition is paramount since AF-related cardioembolic strokes are especially severe and oral anticoagulation (OAC) can prevent recurrent events.¹ Optimal AF detection requires implantable loop recorders (ILR), which are resource-intensive with limited uptake.¹ OAC after stroke without confirmed AF increases bleeding without improving outcomes. Efficient methods are needed to identify AF-related stroke, to guide intensive rhythm monitoring and OAC prescription.

Validated AF risk scores, such as Cohorts for Heart and Aging Research in Genomic Epidemiology-AF (CHARGE-AF),² can discriminate cardioembolic stroke.¹ However, scores are rarely employed due to cumbersome calculation and potential misclassification. Overlapping risk factors between cardioembolic and non-cardioembolic stroke also limits discriminative value.

ECG-AI is a validated deep learning model to estimate AF risk using the 12-lead electrocardiogram (ECG),³ an inexpensive test routinely performed after stroke. Since ECG-AI provides predictive utility beyond clinical risk,³ we hypothesized ECG-AI may be well-suited to discriminate cardioembolism at the time of stroke, and could provide a low-cost method for ILR prioritization.

We retrospectively analyzed 1,597 ischemic stroke patients from Massachusetts General Hospital and Brigham and Women’s Hospital undergoing prospective neurologist-adjudicated stroke subtyping using Trial of ORG 10172 in Acute Stroke Treatment (TOAST) or Causative Classification of Stroke (CCS) criteria.¹ The Mass General Brigham Institutional Review Board approved the study. Informed consent was obtained or waived via protocol-specific allowance. Study data are available upon request.

We included patients with complete CHARGE-AF scores, ≥1 digital-format ECG obtained within 1 year before to 30 days after stroke, and no AF on the ECG used for analysis. Cardioembolism was defined using criteria favoring specificity (TOAST “definite cardioembolism” or CCS “evident or probable cardio-aortic embolism”), and adjudicators were unaware of ECG-AI results.¹ We compared CHARGE-AF, ECG-AI and CH-AI (ECG-AI and CHARGE-AF) for discrimination of cardioembolism from other stroke subtypes using the area under the receiver operating characteristic curve (AUROC) and logistic regression. We assessed associations between ECG-AI and incident AF, defined using an electronic health record-based algorithm (positive predictive value 92%).⁴ In secondary analyses, we included ECGs demonstrating AF, and excluded individuals with AF diagnosed before or 30 days after stroke.

We analyzed 964 individuals with ischemic stroke (age 65±15, 40% women), of which 188 (19.5%) were cardioembolic. ECGs were performed median 1 (quartile-1: 0, quartile-3: 2) day from stroke. ECG-AI (AUROC 0.753 [95%CI 0.712–0.794]) and CH-AI (0.761 [0.722–0.800]) had greater discrimination of cardioembolic stroke than CHARGE-AF (0.675 [0.634–0.716], p<0.01 for both). Odds ratios for cardioembolism were larger for ECG-AI (2.72 per 1-SD increase [2.29–3.27]) and CH-AI (2.97 [2.44–3.65]) versus CHARGE-AF (2.00 [1.65–2.44]). ECG-AI stratified risk of longitudinal AF after stroke (10-year AF risk in top 10% of ECG-AI 60.6% [37.2–75.3], hazard ratio 2.50 [1.59–3.93] versus remainder; 10-year AF risk in bottom 10% of ECG-AI 14.4% [2.5–24.9], 0.34 [0.15–0.77] versus remainder) (Figure). ECG-AI remained favorable to CHARGE-AF when ECGs showing AF were included (n=1,069; AUROC ECG-AI 0.816 [0.785–0.848]; CHARGE-AF 0.683 [0.647–0.718]), and when all individuals with AF before or within 30 days after stroke were excluded (n=733; ECG-AI 0.641 [0.556–0.726]; CHARGE-AF 0.597 [0.504–0.690]).

Figure.

Summary of analysis and results. Panel A depicts an overview of the current study. A total of 964 individuals with acute stroke and paired ECG were included. A total of three AF risk estimation schemes were compared: a) CHARGE-AF, a validated clinical risk score, b) ECG-AI, a deep learning model using 12-lead ECG, and c) CH-AI a combination of CHARGE-AF and ECG-AI. Panel B depicts discrimination performance for each model, using univariable odds ratios on the left and receiver operating characteristic curves on the right. Panel C depicts the cumulative risk of AF among 606 individuals with no AF before or within 30 days after stroke and longitudinal follow-up available after stroke. The left panel depicts the cumulative risk of AF stratified by an ECG-AI score above the 90^th percentile, and the right panel depicts the cumulative risk of AF stratified by an ECG-AI score below the 10^th percentile. Shaded regions denote 95% confidence intervals.

Among ~1,000 acute stroke patients with neurologist-adjudicated stroke subtyping, we compared a validated AF risk score, ECG-based AI risk stratification tool, and their combination. AI-based models discriminated cardioembolic stroke favorably compared to clinical risk factors, and stratified risk of future AF.

Findings yield several implications. First, AI-enabled ECG analysis discriminates cardioembolic stroke. Although ECG-AI discrimination was favorable to CHARGE-AF, AUROC remained moderate and likely insufficient in isolation to classify cardioembolic stroke. Second, integrating clinical risk factors with ECG-based AI adds little value for identifying cardioembolism, likely due to common risk factors for cardioembolic and non-cardioembolic strokes. Third, ECG-AI may improve efficiency of ILR utilization after stroke. Nearly two-thirds of individuals with ECG-AI scores in the top 10% developed new AF within 10 years, identifying a population where ILR may be high-yield. Conversely, 10-year AF risk was 15% in the bottom 10% of ECG-AI, suggesting non-AF etiologies could be prioritized.

Individuals used to train ECG-AI were excluded, but patients are from the same healthcare system. Discrimination was lower when prevalent AF was excluded. Requiring digitized ECGs and complete CHARGE-AF scores introduces bias. Like ECG-AI, CHARGE-AF was designed to estimate AF risk rather than discriminate cardioembolic stroke. Even stroke neurologist-adjudicated subtyping is subject to misclassification. Incident AF detection after stroke depends on monitoring, which was left to the discretion of the clinical team. Monitor utilization is appreciable in our post-stroke population (6-month incidence 20–70% based on subtype⁵).

An AI-enabled AF risk model can discriminate cardioembolic stroke using one 12-lead ECG, performing favorably to a validated 11-component risk factor score. Future work is warranted to determine whether AI models specifically designed to identify cardioembolic stroke can achieve greater discrimination, and whether prospective application of ECG-AI can assist stroke subtype classification and prioritization of intensive rhythm monitoring after stroke.