Machine Learning to Predict Microvascular Obstruction and Left Ventricular Dysfunction in ST-Elevation Myocardial Infarction

Purpose:
To evaluate the predictive accuracy of machine learning models to predict microvascular obstruction (MVO) at 3 days post-infarction, and impaired left ventricular ejection fraction (LVEF) at 3 months in patients with acute ST-elevation myocardial infarction (STEMI).
Methods:
A retrospective analysis was conducted in 200 patients with anterior STEMI from the European Intracoronary Cooling Evaluation (EURO-ICE) trial. Clinical and demographic data were employed to predict MVO and LVEF. Five machine learning models were assessed: Regularized Logistic Regression, Decision Tree, Explainable Boosting Machine, Random Forest, and CatBoost. Predictive accuracy was evaluated using AUC with 5-fold cross-validation with confidence intervals from DeLong’s method.
Results:
For MVO prediction, the machine learning models demonstrated AUCs ranging from 0.651 (Decision Tree [95% CI: 0.58-0.721]) to 0.799 (Random Forest [95% CI: 0.735-0.863]). Logistic Regression achieved 0.723 [95% CI: 0.648-0.797], Explainable Boosting Machine
0.777 [95% CI: 0.709-0.845], and CatBoost 0.789 [95% CI: 0.724-0.854] (Figure). In predicting LVEF, AUCs varied between 0.638 (Decision Tree [95% CI: 0.567-0.709]) and 0.731 (Random Forest [95% CI: 0.658-0.804]). Logistic Regression achieved 0.671 [95% CI:
0.592-0.75], Explainable Boosting Machine 0.707 [95% CI: 0.63-0.784], and CatBoost 0.725 [95% CI: 0.65-0.801].
Conclusion:
Machine learning models demonstrate moderate predictive performance for MVO and LVEF in STEMI patients, with MVO predictions being more accurate across all models. More complex models such as CatBoost and Random Forest outperformed simpler algorithms, highlighting their potential to enhance targeted patient care. Further validation, for instance,
with more external datasets, would further confirm these models' usefulness in clinical settings.