Machine Learning Approach for Motion Artifact Detection in Ballistocardiogram Signals.

With the current increase in cardiovascular disease and the complexities they create, especially for aging seniors, we are working on in-home and non-invasive techniques to monitor vital signs for early detection of health conditions. Ballistocardiography has shown to be useful for long-term evaluation of myocardial strength. We have previously reported the successful utilization of our hydraulic bed sensor in the estimation of heart rate, sleep posture, and blood pressure. However, bed sensors used in naturalistic settings such as the home are known to be highly susceptible to motion artifacts.In this paper, the state-of-the-art methods for motion artifact detection and reduction are reviewed, and a new sequential machine learning approach is proposed. The proposed method is based on 53 novel features extracted jointly from time and frequency domains for noise detection. Our experiments show detection accuracy and sensitivities as high as 99%. Data were collected in two separate IRB approved data collections, one with 16-minute sequences from 25 subjects in the lab and the other with 5 sets of overnight data collected at a sleep center.