Predicting Intramyocardial Hemorrhage Before Reperfusion in STEMI Patients With Intrinsically Explainable Artificial Intelligence

Figures

• 

  Figure 1. Study Flow for the MIRON-PREDICT Cohort Consecutive patients presenting with index ST-segment elevation myocardial infarction (STEMI) were screened for eligibility (n = 324). After exclusion at presentation (n = 20), 304 patients were prospectively enrolled and treated with primary percutaneous coronary intervention (PCI). A further 16 patients were excluded before cardiac magnetic resonance imaging (MRI), yielding a final study cohort of 288 patients with complete clinical data and T2∗ CMR imaging performed 48 to 72 hours after PCI. The final cohort was divided into a development cohort (n = 252; patients enrolled before model development) and an external validation cohort (n = 36; patients enrolled after model development). CMR = cardiac magnetic resonance; eGFR = estimated glomerular filtration rate; PVCs = premature ventricular contractions.

• 

  Figure 2. Superposable Neural Networks Model Architecture and Results (A) Superposable Neural Network (SNN) architecture. The deployed inference model is a compact additive network; training complexity reflects the optimization procedure, including teacher-student distillation, rather than the final model. (B) Relative importance of 12 candidate predictors. Coronary collaterals and the interaction between summed ST-segment elevation and degree of occlusion were sufficient for the final model. (C, D) Learned contribution functions showing how these predictors relate to intramyocardial hemorrhage (IMH) risk. (E) ROC analysis of the SNN. The operating threshold was set at 0.63, corresponding to the point closest to (0,1). Decision rules extracted from the additive functions were then translated into the point-based bedside score shown in Figure 3. AUC = area under the receiver operating characteristic; BMI = body mass index; DM = diabetes mellitus; HTN = hypertension; ROC = receiver operating characteristic.

• 

  Figure 3. Superposable Neural Network-Derived Bedside Intramyocardial Hemorrhage Score Point-based bedside score derived from the final SNN for intraprocedural prediction of intramyocardial hemorrhage (IMH). Points are assigned according to summed ST-segment elevation, degree of culprit-vessel occlusion, and collateral status. The total score classifies patients as lower or higher risk for IMH, with a prespecified decision threshold of 4 or greater. This score preserves the decision logic of the additive SNN while simplifying application during emergency coronary angiography before reperfusion.

• 

  Figure 4. Prospective Model Validation: Intramyocardial Hemorrhage and Non-Intramyocardial Hemorrhage Cases Representative non-IMH and IMH cases from the study. The first column shows pre-PCI ECG examples and summed ST-segment elevation. The second column shows coronary angiography, including degree of occlusion and collateral status. The third column shows the bedside score, with scores < 4 classified as non-IMH and scores 4 or greater classified as IMH. The fourth and fifth columns show CMR confirmation of infarction and intramyocardial hemorrhage using late gadolinium enhancement and T2∗ imaging, respectively. Red contours mark the endocardium, green contours mark the epicardium, and arrows indicate microvascular obstruction or IMH where present. CAG = coronary angiography; IMH = intramyocardial hemorrhage; LGE = late gadolinium enhancement.

• 

  Figure 5. Relationship Between Predictive Scores and Intramyocardial Hemorrhage Volume This figure illustrates the mean intramyocardial hemorrhage (IMH) volume, along with standard error, from patients with hemorrhagic MI stratified by Predictive Score groups from our proposed model. Patients with a Predictive Score of 4 or higher developed IMH with volume as %LV determined using T2∗ CMR. A clear positive relationship was observed, where higher predictive scores were associated with greater IMH volume. LV = left ventricular.

• 

  Central Illustration. Predicting Intramyocardial Hemorrhage Before Reperfusion in ST-Elevation Myocardial Infarction Patients with Intrinsically Explainable Artificial Intelligence ECG = electrocardiogram; PCI = percutaneous coronary intervention; SNN = Superposable Neural Network; STEMI = ST-segment elevation myocardial infarction; XAI = explainable artificial intelligence.