0326 Unobtrusive Detection of Heart-rate-dipping During Sleep Based on Force Sensors in a Smart Bed

Abstract Introduction Typical heart rate (HR) during sleep exhibits a dipping pattern whose absence is associated with increased risk of all-cause mortality and cardiovascular events. Unobtrusive and longitudinal monitoring of HR patterns may be valuable for early detection of cardiovascular risk. We developed a high-temporal coverage algorithm to estimate instantaneous HR enabling characterization of HR patterns using data from a smart bed. Methods Seventeen smart beds were equipped with force sensors under each leg. The force signals acquired at 250 Hz from the sensors located at the bed’s top on the user side were filtered in the 0.5 to 40 Hz band and subsampled to 100 Hz. The data from 226 sleep sessions from 9 participants (5M/3F), mean age 41.7 (SD: 11.3) years-old, were used to train a deep neural network to estimate instantaneous HR every second using data from a 15-second rolling window. The reference HR value was obtained from a Zephyr BioHarness device. The data from 33 sleep sessions from 8 subjects (6M/2F), mean age 41.0 (SD: 11.8) years-old were used to estimate HR, the lowest HR and the latency thereof. The reference HR was obtained from an in-home PSG device (Cerebra). A technician performed the sleep staging and event annotation. The accuracy of estimating instantaneous HR was quantified using a Bland-Altman (BA) analysis at the epoch (30 seconds) level. We defined temporal coverage as the fraction of epochs with HR estimation accuracy ≥ 90%. The HR dipping value was obtained from a smoothed version of the instantaneous HR curve. The HR dipping value and its latency for the ground-truth and the estimated instantaneous HR were compared using BA analysis. Results The limits-of-agreement (LoA) for instantaneous HR estimation were -7.31 to +7.74 beats/min. The mean coverage across sleep sessions was +92.6 %. The LoA for the minimum HR estimation were -1.48 to +3.09 beats/min. The LoA for the minimum HR latency were -2.28 to +1.79 minute. Conclusion Force sensors under a smart bed legs enable accurate and high-temporal coverage HR estimation. This was tested in an HR dipping estimation application resulting in a tight LoA of 4.57 beats/min. Support (if any)