Denoising of Heart Sound Recordings Standard vs. supervised techniques: is cleaner better for classification?

Heart sound recordings are often corrupted by broadband noise which can prevent subsequent knowledgeable analysis. Dedicated denoising techniques have proved effective in significantly increasing the signal-to-noise ratio of heart sounds. In this study, we examine how the effects of both standard and supervised denoising methods translates in terms of automatic classification of healthy and pathological heart sounds. We show that, while supervised denoising can improve classification performance with a sensitivity gain of up to 90%, attempts to clean up heart sounds with standard approaches can also, under several noise conditions, be counterproductive and lead to lower classification scores.Clinical relevance- This study shows that the use of conventional denoising techniques to clean up heart sounds can lead to serious misinterpretation of the normal/abnormal nature of the signals, and that signal processing oriented indicators, such as the signal-to-noise ratio, are not trustworthy in themselves for informed analysis. The denoising procedure must therefore be used with great care, as cleaner signals are not necessarily better if they do not contribute to improve the reliability of medical diagnosis.