AHA Scientific Sessions 2026 Abstract · Nov 2, 2026
Complementary Predictive Utility of Disease-Specific Electrocardiogram-Based Artificial Intelligence Models
By Shahid A, Kany S, Friedman SF et al.
Background: Electrocardiogram-based artificial intelligence (ECG-AI) models can predict future incidence of specific cardiovascular conditions, but the extent to which ECG-AI models encode truly disease-specific signals versus general cardiovascular risk remains poorly understood. If present, the ability to leverage complementary information across distinct models may improve cardiovascular risk stratification using ECG-AI. Research Question: Does concordant high-risk classification across multiple disease-specific ECG-AI models offer complementary risk stratification? Methods: We estimated the risk of incident AF, HF, and stroke across two hospital-based cohorts, Mass General Brigham (MGB) and Beth Israel Deaconess Medical Center (BIDMC), both independent of model development. Three validated ECG-AI models were evaluated: ECG2AF, ECG2HF, and ECG2Stroke. Outcomes were defined over a 10-year horizon using previously validated diagnostic codes, with HF adjudication assisted by a validated EHR-based natural language processing model. Within each cohort, we serially assessed the Kaplan-Meier cumulative risk of each outcome (i.e., incident AF, HF, and stroke) across risk strata defined as follows, with high-risk defined by the top tertile of model score: not high-risk according to disease-specific model (e.g., not high-risk according to ECG2HF for the HF outcome), high-risk according to disease-specific model only, high-risk according to disease-specific model + one additional model, and high-risk according to all three models (Figure 1). Spearman’s rank correlation was quantified to assess shared information across models. Results: Sample sizes and event counts are shown in Figure 1. Spearman’s ρ demonstrated moderate-to-strong correlation among the three ECG-AI models, ranging from 0.76 to 0.82 in MGB and 0.79 to 0.87 in BIDMC (Figure 2). ECG2HF and ECG2Stroke were the most strongly correlated. Across outcomes and cohorts, individuals classified as high-risk by all three ECG-AI models consistently had the highest cumulative risk of incident disease. For instance, 10-year cumulative risk of AF was ~5-fold higher among individuals classified as high-risk by all three models (MGB: 20.4% [95% CI: 19.1-21.8], BIDMC: 23.8% [22.4-25.1]) versus those not classified as high risk by the AF-specific model, ECG2AF (MGB: 3.9% [3.6-4.2], BIDMC: 4.2% [3.9-4.6]), and those classified as high risk by ECG2AF only having an intermediate incidence (MGB: 8.5% [6.7-10.2], BIDMC: 11.7% [9.9-13.5]). A similar gradient of disease incidence was observed for 10-year cumulative risk of HF and stroke (Figure 3). Conclusion: Despite substantial correlation, combining distinct disease-specific ECG-AI model risk signals improves risk stratification. Future work is warranted to define optimal approaches to leverage complementary information contained within distinct AI models.