Armaghane Danesh Bimonthly Journal (Dec 2023)
Improving the Diagnosis of Cardiac Abnormalities Through Feature Extraction from the Heart Sound Signal Using Machine Learning Classification Algorithms
Abstract
Background & aim: Extracting information from the heart sound signal and detecting the abnormal signal in the early stage can play a vital role in reducing the death rate caused by cardiovascular diseases. Therefore, many researches have been done in processing these signals up to now. So, this study aimed to determine the improvement of heart abnormalities diagnosis by extracting features from the heart sound signal by applying machine learning classification algorithms. Methods: The present descriptive–analytical study was conducted at Kazerun Azad University in 2023. The research data were selected from the 2016 Physionet Challenge database. After pre-processing and noise removal, 6 new features and 35 features (41 features) used in previous researches were extracted from the heart sound signals. The 6 new features are " Relative Average Perturbation", " five-point Period Perturbation Quotient", "local shimmer (in dB)", " three-point Amplitude Perturbation Quotient " and " five-point Amplitude Perturbation Quotient " and " correlation of time center of signal and frequency center of signal". The extracted features were applied as input to four classifiers of random forest, support vector machine, K nearest neighbor and linear discriminant analysis. Accuracy, sensitivity and specificity of each classification were calculated. In order to investigate the impact of new features in the diagnosis of cardiac abnormalities, the results obtained were compared with studies that used similar data and classifications but extracted fewer features from the data. The collected data were analyzed using t-tests and logistic regression. Results: The highest accuracy and sensitivity were obtained in the Linear Discriminant Analysis classifier, which are 91.52 and 96.19, respectively. The highest specificity was obtained in the Random Forest classifier at the rate of 88.90. According to the obtained results, by adding new features, the three indices of accuracy, sensitivity and specificity are improved in the two classifiers of K-nearest neighbor and Linear Discriminant Analysis. Extraction of these features also increases the level of specificity in the Random Forest classification. Conclusion: The results indicated that the extraction of new features led to increase in the accuracy, sensitivity and specificity in the diagnosis of cardiac abnormalities compared to the results of previous researches.
