IU AI-Derived Score Predicts Heart-Attack Bleeding Risk

The final analysis included 288 patients from the MIRON-PREDICT study: 252 in the development cohort and 36 in a prospective external validation cohort. Cardiac MRI performed 48 to 72 hours after percutaneous coronary intervention served as the reference for whether intramyocardial hemorrhage occurred; it was identified in 142 patients.

Researchers trained a Superposable Neural Network, an additive architecture designed to expose how each input contributes to a prediction. The model narrowed the decision to three measurements available before reperfusion: summed ST-segment elevation on the electrocardiogram, whether the culprit artery was totally occluded, and whether coronary collateral vessels were present.

Those relationships were converted into a bedside score from zero to six points. A score of four or more classified a patient as higher risk. In the development cohort, the point score achieved 84.9% accuracy, 82.3% sensitivity, and 87.3% specificity. Applied without refitting to the 36-patient external cohort, it achieved 83.3% accuracy, 81.3% sensitivity, and 85% specificity.

The external cohort is useful because it was enrolled after model development and evaluated with the prespecified score. It is also small. The authors explicitly call for validation in larger, multicenter cohorts to test whether performance generalizes across populations and clinical settings.

The reported metrics measure classification against later MRI findings; they do not show that using the score changes treatment or improves patient outcomes. For clinical-AI teams, the notable design choice is the translation from an interpretable neural model into a compact rule that can be calculated from existing catheterization-lab data. Prospective impact studies and workflow validation would still be needed before broad clinical adoption.