Key Techniques and Challenges for Processing of Heart Sound Signals

Recently, new advances and emerging technologies in healthcare and medicine have been growing rapidly, allowing for automatic disease diagnosis. Healthcare technology advances entail monitoring devices and processing signals. Advanced signal processing and analytical techniques were effectively implemented in numerous research domains. Thus, adopting such methods for biomedical signal processing is an essential study field. The signal processing techniques are explicitly applied to heart sound (called phonocardiogram or PCG) signals as part of biomedical signals for heart health monitoring in this paper. The automatic detection of life-threatening cardiac arrhythmias has been a subject of interest for many decades. However, the computer-based PCG segmentation and classification methods are still not an end-to-end task; the process involves several tasks and challenges to overcome. The conducted evaluation scheme of the classifier also has a significant impact on the reliability of the proposed method. Our main contributions are twofold. First, we provided a systematic overview of various methods that can be employed in real applications for heart sound abnormalities. Second, we indicated potential future research opportunities. PCG segmentation is critical, and arguably the hardest stage in PCG processing. Basically, basic heart sounds can be identified by detecting the offset R-peak and T-wave in the ECG signal. Unfortunately, utilizing the ECG signal as a reference to the PCG segment is not always an easy operation because: it requires synchronous recording of ECG and PCG signals; precise identification of T-wave offset is often difficult; and ECG-PCG temporal alignment is not always consistent. Using machine learning methods in PCG segmentation involves multiple types and many features retrieved in both univariate or multivariate formats. This leads to selecting the best PCG-segmentation performance feature sets. PCG segmentation approaches that use featureless methods based on powerful statistical models have the potential to solve the problem of feature extraction and minimize the total computational cost of the segmentation approach.