A hybrid model for recognizing cardiac murmurs from phonocardiogram signal

Heart murmurs are the sounds heard in addition to lub (S1) and dub (S2) due to valve dysfunction, septal defect and other abnormalities of heart which can be fatal. In this paper, a novel hybrid model involving machine learning and signal processing techniques is proposed for the diagnosis of heart sounds using Phonocardiogram (PCG) signals. A dataset comprising of 175 normal and 91 abnormal PCG signals is used. The authors propose a modular approach comprising of four levels of classification supported by preprocessing and segmentation. Preprocessing employs minimum statistics approach to suppress noise and enhance signal quality. At every stage of classification, the signal is characterized by parameters from its time-frequency, spectral and statistical analyses. In the first level, normal heart sounds are separated from the abnormal, using a Support Vector Machine (SVM) classifier having 91.11% accuracy. In the second level, an abnormal signal is identified as a valvular or nonvalvular murmur with 93.70% precision by employing SVM. The properties of Shannon energy envelogram of PCG signal is used in segmentation process to obtain systole and diastole segments. In the third level, these segments are analyzed for systolic and diastolic murmurs by thresholding, the accuracy being 90.61%. In the final level, systolic murmurs are segregated across Aortic Stenosis and Mitral Regurgitation with 83.89% accuracy by using SVM and diastolic murmurs are identified as Mitral Stenosis or Aortic Regurgitation with 100% accuracy by thresholding approach. Sensitivity and Specificity are used as performance metrics to achieve a robust and competent system.