An explainable artificial intelligence framework using electrocardiograms to predict mortality and major adverse cardiovascular events

Artificial intelligence (AI) algorithms can detect subtle patterns in electrocardiograms (ECGs) to predict cardiovascular risk. Though accurate, these models are often unable to explain the factors driving their predictions, limiting interpretability and clinical adoption. We developed CLAIRE, an explainable AI framework leveraging an open-source chain-of-thought large language model (DeepSeek-R1-0528-Qwen3-8B) and structured ECG parameters to forecast mortality and major adverse cardiovascular events (MACE). Using 60,000 adult ECGs, the framework consisted of two stages: CLAIRE-α, trained on 636 unique ECG features, age, and gender to predict outcomes, and CLAIRE-β, which identified the most informative features and generated physiologically plausible explanatory pathways linking ECG abnormalities to clinical endpoints. Feature-reduced CLAIRE-α variants were subsequently trained using top-ranked features from CLAIRE-β. Models using all 636 features achieved the highest performance for MACE (accuracy/AUROC 0.97/0.98) and mortality (accuracy/AUROC 0.86/0.86) prediction. Feature-reduced variants retained good to excellent discrimination while lowering computational burden. Model-generated explanations highlighted relationships consistent with plausible cardiovascular physiology, as assessed by board-certified physicians. CLAIRE combined predictive accuracy with interpretability by identifying physiologically meaningful ECG features and providing hypothesis-generating explanations for its risk predictions. This study represents a step toward transparency and addresses a key limitation of contemporary ECG-based AI models.