Automatická detekce infarktu myokardu z EKG záznamů s využitím analýzy ST segmentů

Abstract

This thesis focuses on automatic detection of myocardial infarction (MI) from 12-lead ECG using machine learning (ML) methods. In addition, the work was extended to localize the three most common types of MI: inferior, lateral and anterior wall. Interpretable ML methods were selected for MI detection and localization: KNN, LDA, SVM and ensemble methods based on decision trees (RF, ET, XGBoost, CatBoost). Individual MI detectors and localizers were trained on features extracted from 8201 different ECG recordings from the PTB-XL database for each of the 12 leads. Features associated with the ST segment were primarily extracted. Subsequently, after performing a search, other features were found to reflect MI. Therefore, the feature space was expanded to include additional attributes not necessarily related to the ST segment. This was followed by objective feature selection using advanced ML methods: recursive features elimination (RFE), Fisher score. According to the results of MI detection and localization, the RFE selection method provided more relevant attributes than attributes filtering based on Fisher score. When features were selected for MI detection, the RFE method mainly related to the ST segment. On the other hand, for MI localization, features related to the QRS complex were selected by RFE method. The most successful MI detector is considered to be the XGBoost-based detector detecting MI on a reduced set of features by RFE with an accuracy of 86.71 %±0.90 %. In the case of MI localization, the CatBoost-based model generalized best to the non-reduced features set with an accuracy of 78.25 %±2.34 %. Moreover, this work is the first to use interpretable ML methods to train MI detectors and localizers on the PTB-XL database.