Prediction of left main coronary artery lesion in patients with stable coronary artery disease using artificial intelligence technology

Aim. To develop a software package that predicts the presence of significant stenosis of the left main coronary artery (LMCA) in patients with stable coronary artery disease (CАD) based on clinical and instrumental signs using machine learning (ML) methods. Material and methods. The study included 208 patients with stable CAD, hospitalized for planned invasive coronary angiography (ICA). Based on the results of ICA, two groups of patients were identified: the main group - 104 patients with hemodynamically significant stenosis of the LMCA and the control group - 104 patients without obstructive lesions of the coronary arteries. Stenosis of the LMCA of more than 50% was considered significant. Based on the study of case histories, an Excel database was created containing information on 107 clinical and instrumental features of all patients. Based on the results of a preliminary significance analysis, 63 potentially important features were selected for further use in the classifier model. Then, 59 of the most important features from the point of view of ML were identified. Given the nature of the features under consideration, gradient boosting was chosen as the ML algorithm. The programs developed during the study were written in the Python programming language in the PyCharm integrated programming environment. Results. Based on the most informative features, models for predicting LMCA damage in patients with CAD were constructed. Gradient boosting was determined as the optimal ML algorithm. Three of its implementations were considered: LightGBM (Light Gradient Boosting Machine), XGBoost (Extreme Gradient Boosting) and CatBoost (Categorical Boosting). The CatBoost model turned out to be the most effective for classifying the presented objects. The software package using the CatBoost model demonstrated the accuracy of the classifier prediction of 86.2% on the test data set with a sensitivity and specificity of 77% and 76.9%, respectively. According to the constructed learning curves, it was found that further expansion of the training data volume will improve the quality of the model. Conclusion. The developed software package for identifying hemodynamically significant stenosis of the LMCA in patients with CAD has good prognostic accuracy with the prospect of further training. The obtained solution can be integrated into the diagnostic process as part of an application for a personal computer or a web interface in order to provide support for medical decision-making.