Heart sound classification from unsegmented phonocardiograms

Abstract

Objective: Most algorithms for automated analysis of phonocardiograms (PCG) require segmentation of the signal into the characteristic heart sounds. The aim was to assess the feasibility for accurate classification of heart sounds on short, unsegmented recordings. Approach: PCG segments of 5 s duration from the PhysioNet/Computing in Cardiology Challenge database were analysed. Initially the 5 s segment at the start of each recording (seg 1) was analysed. Segments were zero-mean but otherwise had no pre-processing or segmentation. Normalised spectral amplitude was determined by fast Fourier transform and wavelet entropy by wavelet analysis. For each of these a simple single feature threshold-based classifier was implemented and the frequency/scale and thresholds for optimum classification accuracy determined. The analysis was then repeated using relatively noise free 5 s segments (seg 2) of each recording. Spectral amplitude and wavelet entropy features were then combined in a classification tree. Main results: There were significant differences between normal and abnormal recordings for both wavelet entropy and spectral amplitude across scales and frequency. In the wavelet domain the differences between groups were greatest at highest frequencies (wavelet scale 1, pseudo frequency 1 kHz) whereas in the frequency domain the differences were greatest at low frequencies (12 Hz). Abnormal recordings had significantly reduced high frequency wavelet entropy: (Median (interquartile range)) 6.63 (2.42) versus 8.36 (1.91), p < 0.0001, suggesting the presence of discrete high frequency components in these recordings. Abnormal recordings exhibited significantly greater low frequency (12 Hz) spectral amplitude: 0.24 (0.22) versus 0.09 (0.15), p < 0.0001. Classification accuracy (mean of specificity and sensitivity) was greatest for wavelet entropy: 76% (specificity 54%, sensitivity 98%) versus 70% (specificity 65%, sensitivity 75%) and was further improved by selecting the lowest noise segment (seg 2): 80% (specificity 65%, sensitivity 94%) versus 71% (specificity 63%, sensitivity 79%). Classification tree with combined features gave accuracy 79% (specificity 80%, sensitivity 77%). Significance: The feasibility of accurate classification without segmentation of the characteristic heart sounds has been demonstrated. Classification accuracy is comparable to other algorithms but achieved without the complexity of segmentation.