For heart attack patients rushing toward emergency intervention, the damage inflicted by restoring blood flow can paradoxically worsen outcomes — a phenomenon linked to bleeding within the heart muscle itself. Until now, clinicians had no reliable way to identify which patients faced this hidden complication before treatment began. A new explainable AI framework may change that calculus meaningfully for high-risk STEMI care.

Drawing on 288 STEMI patients enrolled in the prospective MIRON-PREDICT clinical study between mid-2023 and late 2024, researchers developed a neural network — a Superposable Neural Network designed for interpretability — to predict intramyocardial hemorrhage (IMH) before reperfusion is attempted. IMH was confirmed via cardiac MRI at 48–72 hours post-intervention in 142 of 288 patients, reflecting a 49% incidence in this cohort. The final predictive model distilled three variables from electrocardiographic, angiographic, and clinical data: presence of coronary collateral vessels, degree of coronary artery occlusion, and sum ST-score from ECG. Validated on a held-out set of 36 patients, the model achieved 84.9% accuracy, 82.3% sensitivity, and 87.3% specificity — performance metrics that, if replicated, would represent a clinically actionable advance. Crucially, these inputs were translated into a point-based bedside scoring tool requiring no specialized imaging.

This work sits at an important intersection: IMH affects roughly 40% of reperfused STEMI patients and independently predicts adverse remodeling and mortality, yet it has remained essentially undetectable pre-procedure. The explainability architecture matters here — cardiology's resistance to opaque AI is well documented, and a transparent model tied to mechanistically plausible variables (collateral flow protects myocardium; total occlusion and high ST-burden signal greater ischemic insult) is far more likely to achieve clinical adoption. Key limitations include a relatively modest validation cohort of 36 patients, single-study derivation, and the absence of outcome data linking model-guided decisions to improved survival. This is an incremental but structurally sound proof-of-concept that warrants multicenter prospective validation.