Classification of heart sounds using time-frequency method and artificial neural networks

Digitally recorded pathological and non-pathological phonocardiograms (PCGs) were characterised by a time-frequency (TF) method known as the trimmed mean spectrogram (TMS). Features were extracted from the TMS containing the distribution of the systolic and diastolic signatures in the TF domain. Together with the acoustic intensities in systole and diastole, these features were used as inputs to the probabilistic neural networks (PNNs) for classification. A total of 56 PCGs were employed to train the PNNs including 21 non-pathological and 35 pathological PCGs. The PNNs were then tested with a different group of 18 non-pathological and 37 pathological PCGs. The system provided a sensitivity of 97.3% (36/37) and a specificity of 94.4% (17/18) in detecting pathological systolic murmurs. The results show that the system offers a promising methodology for classifying murmurs.