Detekce infarktu myokardu z elektrokardiografických záznamů s využitím analýzy frekvenční domény

Abstract

Cardiovascular diseases are the leading cause of death worldwide, and myocardial infarction (MI) is one of its most severe forms. Early diagnosis plays a crucial role in MI, as irreversible damage to the hearttissueoccurswithin20–40minutesofanarteryocclusion. Thisthesisfocusesontheautomatic detection of myocardial infarction from 12-lead electrocardiographic recordings using frequency domain analysis. Recordings from 160 patients from the PTB database are first preprocessed and then decomposed into six frequency bands using the stationary wavelet transform (SWT). The power spectral density (PSD) is estimated for each band using Welch’s method, from which 16 features are extracted, totaling 1152 features per patient record. The number of features is further reduced by a three-phase algorithm consisting of an ANOVA F-test, the ReliefF algorithm, and a correlation filter. Configurations with a maximum number of features of 10, 15, 20, 30, 40, 50, and 100 are systematically tested. The resulting features are classified using eight machine learning algorithms: SVM, Random Forest, kNN, LDA, Naive Bayes, Voting, Stacking, and AdaBoost. The performance of the classifiers is evaluated using a 10-fold stratified cross-validation, for reduced configurations, a nested cross-validation is applied. The Voting classifier achieved the best results without feature reduction, reaching an accuracy of 98.1 % and an AUC value of 0.995. The results show that when reduced to just 20 features, some classifiers maintain a performance comparable to using the full feature set, which indicates high redundancy in the original feature space.