Simultaneous pulse rate estimation for two individuals that share a sensor-laden bed

Abstract

Sleep monitoring has received increased attention in recent years given an improved understanding of the impact of sleep quality on overall well-being. A Kansas State University team has developed a sensor-based bed that can unobtrusively track sleep quality for an individual by analyzing their ballistocardiograms (BCGs) while they lay on the bed, foregoing the need to visit a sleep clinic to quantify their sleep quality. A BCG is a signal that represents cardiac forces that have spread from the heart to the rest of the body – forces that result in part from the injection of blood into the vascular system. The sensor bed software can extract BCG-based health parameters such as heart rate and respiration rate from data acquired continuously throughout the night. Such a toolset creates a new challenge, namely that many people sleep on a shared bed. In such cases, a given sensor bed would acquire mixed BCGs that contain information for both people. This thesis documents efforts to create an algorithm to extract individual health parameters from mixed parent BCGs obtained from bed sensors that reside on a shared bed. The first component of the two-part algorithm performs ‘blind source separation:’ a technique originally designed for mixed audio applications that attempts to optimally separate two individual BCGs contained in an original mixed signal. The second component of the algorithm utilizes a frequency-domain, peak-scoring method to identify the most likely fundamental BCG harmonic for each separated signal – a harmonic that corresponds to the pulse rate for that individual. The peak-scoring approach allows the algorithm to overcome challenges associated with different time-domain BCG waveform shapes, the presence of signal artifact, and the loss of BCG characteristic features that occurs during the separation stage. These challenges can be problematic for time-domain pulse rate algorithms, but the repetitive waveform patterns can be exploited in the frequency-spectrum. The peak-scoring algorithm was verified by comparing pulse rates determined from single-subject BCGs (obtained in various sleeping positions) against pulse rates determined from simultaneously collected electrocardiograms. The separation and peak-scoring components were combined together, and this overall technique was applied to over 20 sets of paired BCG data, with variations in sensor placement, sensor type and mattress type. Early results indicate the ability of the algorithm to determine pulse rates from mixed BCGs with acceptable levels of success but with areas for improvement.