Prediction of coronary microcirculatory dysfunction in STEMI patients: a machine learning approach (TAMIR study)

Background There is no clinical tool that allows clinicians to screen patients for coronary microvascular dysfunction (CMD). This study aims to create a machine learning model for predicting CMD using common biomarkers and compare its performance to one created using traditional statistical methods. Methods This was a prospective multicenter study with patients > 40 years with STEMI and multivessel disease who underwent primary PCI followed by a staged PCI three months later. Coronary flow reserve, fractional flow reserve and index of microvascular resistance were assessed to establish the prevalence of CMD. Multiple machine learning methods as well as traditional logistic regression were utilized to create a predictive models for CMD based on 17 biomarkers (Figure 1). The models were assessed with 10-fold cross validation and an independent validation set. Results 149 patients were enrolled the training set, and 51 patients were enrolled the validation set. The random forest (RF) model was the best performer with the lowest 10-fold cross-validation. Logistic regression was inferior to the RF model. Trimethylamine N-oxide, platelet aggregation with epinephrine and BNP were the three most influential variables in the RF model (Figure 2). The RF model performed well when applied to the validation set: accuracy = 0.92 (95% CI 0.81-0.98), sensitivity = 0.91, specificity = 1.00, negative predictive value = 0.64, positive predictive value = 1.00 and precision = 1.00. Conclusion A machine learning-based random forest model outperformed traditional logistic regression in predicting CMD, demonstrating high accuracy and robustness in validation. This model offers a promising tool for non-invasive CMD screening in clinical practice.