Long-Term Bowel Sound Monitoring and Segmentation by Wearable Devices and Convolutional Neural Networks.

Abstract

Bowel sounds (BSs), typically generated by the intestinal peristalses, are a significant physiological indicator of the digestive system's health condition. In this study, a wearable BS monitoring system is presented for long-term BS monitoring. The system features a wearable BS sensor that can record BSs for days long and transmit them wirelessly in real-time. With the system, a total of 20 subjects' BS data under the hospital environment were collected. Each subject is recorded for 24 hours. Through manual screening and annotation, from every subject's BS data, 400 segments were extracted, in which half are BS event-contained segments. Thus, a BS dataset that contains 20×400 sound segments is formed. Afterwards, CNNs are introduced for BS segment recognition. Specifically, this study proposes a novel CNN design method that makes it possible to transfer the popular CNN modules in image recognition into the BS segmentation domain. Experimental results show that in holdout evaluation with corrected labels, the designed CNN model achieves a moderate accuracy of 91.8% and the highest sensitivity of 97.0% compared with the similar works. In cross validation with noisy labels, the designed CNN delivers the best generability. By using a CNN visualizing technique-class activation maps, it is found that the designed CNN has learned the effective features of BS events. Finally, the proposed CNN design method is scalable to different sizes of datasets.