Classification of Valvular Diseases using Phonocardiogram without Segmentation and Gammatone Filters

Due to current life style and stress levels a large part of the population has become prone to cardiovascular diseases (CVD), which may be due to malfunctioning of the heart valves or narrowing or blockage of the arteries. CVD may lead to death if not diagnosed and treated on time. Since ages auscultation or heart sound is used as a first screening aid for detection of cardiovascular abnormality. The phonocardiogram (PCG) signal captured by placing a stethoscope on the chest reflects the condition of cardiovascular system(CVS). In the current study features from raw PCG signal, without segmentation are used for classification of cardiac valvular diseases. In other methods without segmentation, features like mel frequency cepstral coefficients(MFCCs), from 20ms window with 10ms overlap were used. Since the average cardiac cycle is 0.6s to 0.8s, window length of 0.8s with an overlap of 0.2s is proposed for feature extraction, which reduces the computational burden. Gammatone filter cepstral coefficients(GFCCs), spectral and time domain features are extracted from each window. A random forest classifier is trained with features extracted from 80% of the data and tested with 20% data. The sensitivity, specificity and accuracy obtained are: 99.39%, 99.85% and 99.76% respectively.