Using time-frequency features to recognize abnormal heart sounds

Disease of the heart accounts for 6% of all death. Heart sound is a routine in physical examination clinically, and is sensitive in detecting a subset of heart diseases. In the current study, we build up a model to classify heart sounds for preclinical screening. Heart sounds are recorded under uncontrolled environment, and each sample can range from 5 to 120 seconds. This is a work raised by annual PhysioNet/CinC Challenge. Because of timing and tonic natures of heart events, we used time-frequency features to classify heart sounds in this study. Firstly, each heard sound recording was segmented into cycles using Springer's improved version of Schmidt's method. Each cardiac cycle was cut into 10 partitions and data points were obtained by zero-padding in each partition. Spectral features were extracted from each partition using fast-Fourier Transform (FFT) thus a 3,500 feature matrix was created. Using filter method, 40 features were selected for the final classifier. The average feature matrix of each cycle was then applied to a classification system using 2-means clustering and artificial neural network (ANN). By clustering the unsure class was recognized. The discrimination of normal and abnormal heart sound were performed by a well-trained ANN model. The results showed that our proposed method got a performance with an accuracy 86.5%, a sensitivity 84.4%, a specificity 86.9%. Here we show that classifying abnormal heart sound is a really difficult task due to the heterogeneity of “abnormal events” and intra-sample deviation.