Diagnosis of cardiovascular disorders by machine learning methods

Auscultation and phonocardiography are heart sound analyzing methods which have been preferred by clinicians for many years. Physicians make diagnosis by listening to the heart sounds but this task depends on the listeners' having sufficient experience and ability. For interns and young physicians, distinguishing and interpreting these sounds with high accuracy is a difficult task. Auscultation, which is based on the listening ability of the human ear, is a subjective matter and the same heart sounds can be interpreted in different ways by different listeners. In addition, computer aided systems are needed to use heart sounds to detect cardiovascular diseases in rural areas, home care units and rural health centers where specialist physicians are not available or it is difficult to reach them. In this thesis, computer aided methods have been developed to help specialist physicians in diagnosing murmurs from heart sounds. Data mining and artificial intelligence-based approaches are used in the proposed methods. For this purpose, firstly, the performances of classical machine learning approaches in classifying normal/abnormal PCG records were examined. These methods were used to compare the performances of the original methods developed. In the second proposed application, a deep network model with convolution and bidirectional long-short-term memory layers is trained with MFCC feature matrices extracted by filtering heart sounds. In the third method, frames were extracted from PCG records by optimizing entropy and energy values with the Artificial Bee Colony algorithm. The spectrograms extracted from these frames are classified by the deep network model with convolution and bidirectional long-short-term memory combination layers. Finally, the bidirectional long-short-term memory model and the convolutional neural network model supported by long-short-term memory layers were compared in classifying normal and arrhythmic ECG recordings. The proposed systems are tested on publicly available datasets and their performance values are compared. The performance results obtained in the experiments with the suggested applications are at the level of competition with other current methods. These developed methods are at a level that can help physicians in decision-making.