Learning Optimal Time-Frequency Representations for Heart Sound: A Comparative Study

Computer audition based methods have increasingly attracted efforts among the community of digital health. In particular, heart sound analysis can provide a non-invasive, real-time, and convenient (anywhere and anytime) solution for preliminary diagnosis and/or long-term monitoring of patients who are suffering from cardiovascular diseases. Nevertheless, extracting excellent time-frequency features from the heart sound is not an easy task. On the one hand, heart sound belongs to audio signals, which may be suitable to be analysed by classic audio/speech techniques. On the other hand, this kind of sound generated by our human body should contain some characteristics of physiological signals. To this end, we propose a comprehensive investigation on time-frequency methods for analysing the heart sound, i.e., short-time Fourier transformation, wavelet transformation, Hilbert-Huang transformation, and Log-Mel transformation. The time-frequency representations will be automatically learnt via pre-trained deep convolutional neural networks. Experimental results show that all the investigated methods can reach a mean accuracy higher than 60.0%. Moreover, we find that wavelet transformation can beat other methods by reaching the highest mean accuracy of 75.1% in recognising normal or abnormal heart sounds.