AI-Assisted MI diagnosis from echocardiogram videos: does explainability enhance human-AI collaborative accuracy?

Echocardiography after myocardial infarction (MI) provides clinically useful information through assessment of regional wall motion abnormalities, but interpretation requires expertise and remains subject to observer variability. Artificial intelligence (AI) shows promise in automatic interpretation, but it is unclear how explainability affects human-AI collaborative performance. A ResNet18-LSTM model was trained to classify normal vs MI on 127 apical four chamber (A4C) and 120 apical two chamber (A2C) echocardiogram videos from the HMC-QU dataset. Gradient-weighted Class Activation Mapping (Grad-CAM provided visual explanations. Eight cardiology trainees compared diagnostic performance across three conditions: (a) echo clips alone, (b) echo clips with AI predictions, and (c) echo clips with AI predictions plus Grad-CAM explanations. The AI models demonstrated strong discriminative performance with AUCs of 0.9429 (A2C) and 0.9250 (A4C). AI alone achieved 80.0% accuracy versus 77.0% for clinicians alone. Surprisingly, combining AI with human judgment did not improve performance, and introducing visual explanations reduced accuracy to 72% and specificity from 93.8% to 83.8% (p = 0.046). While AI models can effectively detect MI on echocardiographic videos, current explainability techniques may misalign with clinical reasoning, potentially impairing diagnostic performance. Future integration requires AI visual explanation strategies that complement clinician expertise. Heart attacks (myocardial infarction) remain a leading cause of death worldwide. Doctors use ultrasound videos of the heart (echocardiograms) to detect damage from heart attacks, but interpreting these videos requires significant expertise and results can vary between doctors. Artificial intelligence (AI) shows promise in helping doctors analyze these videos, but it is unclear whether showing doctors how the AI makes its decisions improves diagnostic accuracy. In this study, we trained an AI system to detect heart attack damage from echocardiogram videos using 247 recordings from two standard heart views. We then tested whether providing visual explanations of the AI’s reasoning (heat maps highlighting important regions) would help eight trainee cardiologists make better diagnoses. The AI achieved 80% accuracy on its own, slightly better than the doctors’ 77% accuracy. Surprisingly, when doctors saw the AI’s predictions, their accuracy did not improve. Even more unexpectedly, adding visual explanations reduced their accuracy to 72% and made them more likely to misidentify healthy hearts as abnormal. Our findings suggest that simply making AI systems transparent does not guarantee better decision-making. Future AI tools for heart imaging will need visual explanations specifically designed to match how doctors think and work, not just show where the AI is looking.