Deep learning-based electrocardiographic differentiation of Takotsubo syndrome from acute anterior myocardial infarction: explainability, human-AI interaction, and a utilization bias hypothesis

Background and Aims: Artificial intelligence (AI) may help differentiate Takotsubo syndrome from acute anterior ST-segment elevation myocardial infarction (STEMI) on electrocardiograms (ECGs). We developed an ECG image-based model and evaluated whether AI assistance and visual explanations improved reader performance, overall and by reader background. Methods: A self-supervised three-architecture deep learning ensemble was developed and externally validated using 12-lead ECG images from two institutions, recorded ≤14 days after onset regardless of reperfusion. A four-phase reader study compared unaided interpretation, a reference sheet, and AI assistance (Phases 1—3). In Phase 4, readers were randomized to the reference sheet alone, AI predictions, or predictions with gradient-weighted class activation mapping (Grad-CAM). Correctness was analyzed with generalized linear mixed-effects models. Results: Cohorts comprised 832 development and 95 external validation participants. External validation accuracy was 0.914, and the area under the receiver operating characteristic curve was 0.970 (95% CI, 0.938—0.992), or 0.958 using first recorded ECGs. Among 70 readers, the estimated probability of correct diagnosis increased from 0.632 unaided to 0.864 with AI (odds ratio [OR], 3.71; P<0.001). In Phase 4, AI assistance improved accuracy vs reference sheet (OR, 4.28; 95% CI, 3.08—5.95), and Grad-CAM added value (OR, 1.53; 95% CI, 1.04—2.25). In exploratory analyses, less experienced readers gained more; with AI, non-cardiologists outperformed cardiologists. Mean AI-assisted reader accuracy remained below model accuracy. Conclusions: An ECG image-based ensemble differentiated Takotsubo syndrome from anterior STEMI and improved reader performance, particularly with visual explanations. Experience-related differences emphasize the need for effective human-AI interaction.