Towards fully automated machine learning and processing of ballistocardiograph signal for heartbeat detection

• Main Author: Zhang, Shiyu
• Other Authors: Lin Zhiping
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Engineering::Electrical and electronic engineering::Electronic systems::Signal processing
• Online Access: https://hdl.handle.net/10356/141320
• Institution: Nanyang Technological University

Cardiovascular diseases are the leading cause of death globally, and ballistocardiogram (BCG) provides a promising and alternative approach to early detection and monitoring of certain pathological signals. While accurate and precise heartbeat detection is the prerequisite for BCG-based clinical solutions, it still remains a signal processing and computing challenge due to its high signal variability and susceptibility to severe artifact corruptions. Recent years have seen a growing interest in using rapidly ad vancing deep learning techniques to overcome the challenge. Existing attempts in this area, however, were primarily focused on classification of per-segmented BCG signals, which are often irrelevant to real application scenarios. Hence, we propose to focus on processing sequential (non-segmented) BCG using appropriate designs of machine learning techniques. In this dissertation, we examine three key techniques including: auto-labeling of samples in a BCG time-series with an ECG reference for building high-quality training data set; deep learning of the association between the BCG time-series and the label time-series; and heartbeat detection in the BCG time-series with predicted sample-by-sample labels. For design and evaluation of these techniques, we built a data set of 16 BCG recordings, from 8 human subjects and 2 conditions: pre-exercise and post-exercise. Experimental results show that our proposed algorithm achieved, on average, high values for coverage (98.73%), heart rate accuracy (98.43%), and low value for heart rate root mean square error (1.64 bpm), which indicate an improvement compared to the algorithm reported in the literature.