A Binarized CNN-Based Bowel Sound Recognition Algorithm With Time-Domain Histogram Features for Wearable Healthcare Systems

Abstract

Human bowel sounds (BSs) convey valuable information on gastrointestinal health. Recently, the application of wearable healthcare systems has made it possible to monitor long-term human BSs. However, for the sake of protecting user privacy and reducing the data transmission volume, it is necessary to conduct coarse-grained BS sifting at the wearable sensor sides. In this brief, an edge BS recognition algorithm is proposed, which aims to pick the resting-state BS events and effectively reject the sound segments which contain only the background noises such as the speech and white Gaussian noise (WGN). The algorithm is highlighted with a time-domain histogram feature and a binarized CNN that is uniquely optimized for BS recognition. Experimental results show that the proposed algorithm reaches 99.92% classification accuracy with a very low false alarm rate. Moreover, the algorithm is validated by hardware implementation and the computation overhead reduction ratio reaches 58.28 in terms of overall operations, in comparison with similar works in literature. These advantages indicate that our algorithm is suitable to be deployed on those edge devices with limited resources.