Automatic Classification of Diseases from Heart Sounds

The number of people who die from cardiovascular diseases is too high to be underestimated. For the diagnosis of these diseases, specialist doctors should listen to the patients with the auscultation method (listening by the specialist physician). However, in this method, which is performed by listening, the skill and experience of the doctor are of great importance and the results may contain subjectivity. For these reasons, it is necessary to develop a suitable algorithm for the diagnosis of heart diseases by digitizing heart sounds in order to classify heart sounds more objectively. Thus, it is expected that the diagnoses made will be more accurate and faster. The proposed study, it is aimed to develop a decision support system that can make an automatic diagnosis to make an objective and high-accuracy diagnosis of heart diseases. In the developed algorithm, discrete wavelet transform (DWT) is applied to heart sound signals from existing data sets. The obtained values: variances, power spectrum density (PSD) values in different frequency ranges, dimensions, the standard deviation of their dimensions, and finally the average values of their dimensions were extracted as features. Then, using these features, disease groups were classified by machine learning methods. According to the system training and test results, the best result was obtained with the “Subspace KNN” classifier and the disease groups were classified with 91.2% accuracy.