Prediction of acute occlusive coronary artery lesion in a contemporary cohort of patients with non-ST-elevation acute coronary syndrome using machine learning

Authors:
Publisher website
Google Scholar
Abstract
Objective. To develop a multivariate predictive model for the early detection of acute coronary occlusion (ACO) in patients with non-ST-elevation acute coronary syndrome (NSTE-ACS) using machine learning methods and a set of predictors available within the first hours of hospitalization. Material and methods. This retrospective observational study included 1,144 patients diagnosed with non-ST-elevation myocardial infarction or unstable angina, hospitalized at a regional vascular center between 2019 and 2021. Invasive coronary angiography was performed in 900 patients, with ACO diagnosed in 64 (7.1%). To identify predictors of ACO, we analyzed 149 clinical, anamnestic, laboratory, and instrumental parameters, available during the initial hours of hospitalization. After selecting predictors, predictive models were developed using machine learning techniques: multivariate logistic regression (MLR), random forest, stochastic gradient boosting (XGBoost), and categorical boosting (CatBoost). Model performance was evaluated based on ROC-AUC, sensitivity (Sens), specificity (Spec), Precision-Recall AUC (PR-AUC), Brier-score, positive predictive value (PPV), negative predictive value (NPV), F1-score, and Accuracy. Results. Independent predictors of ACO included newly diagnosed or new left ventricular regional wall motion abnormalities, erythrocyte sedimentation rate >15.5 mm/h, high-density lipoprotein cholesterol 82 U/L, and signs of ongoing myocardial ischaemia at admission: refractory or recurrent chest pain and/or dyspnea, combined with at least one other established criterion of very high-risk NSTE-ACS according to current clinical guidelines. The CatBoost-based model demonstrated the best prognostic performance (ROC-AUC = 0.841). A highly interpretable prognostic model based on multivariable logistic regression, using the identified risk factors, had performance metrics comparable to the CatBoost model. Conclusion.  The combination of clinical, laboratory, and echocardiographic parameters enables effective prediction of ACO in the NSTE-ACS cohort. The developed machine learning models demonstrate diagnostic accuracy and can form the basis for modifying early risk stratification algorithms and optimizing invasive treatment strategies.

This publication has 28 references indexed in Scilit: