Note:  L. Rosales and B.Y. Su contributed equally to this work.
Abstract: Ballistocardiogram signals produced by hydraulic transducers placed under the bed mattress are used to estimate heart rate using two proposed methods. The first method uses features and clustering extracted in the temporal domain, and the second applies the Hilbert transform and Fourier analysis in the frequency domain. The two methods are evaluated using the data obtained from four senior residents, two with cardiac history, ages 86, 89, 91 and 99. Over five minutes of initial recordings, the minimum and median errors over the four subject data for the clustering method are 0.96% and 5.6%, while those from the Hilbert transform method are 0.59% and 1%. Extensive study of data collected from the subjects acquired over a two to four months period under in-home living conditions showed a median of the percentage agreement of the two methods of 67% with a tolerance of ±3 bpm and 83% with ±6 bpm tolerance. The percentage difference comes from the ability of the two methods in estimating the heart rate under different conditions. Indeed, the two methods are shown to complement each other in heart rate tracking ability and noise resilience, which provides opportunity for fusion in achieving more reliable and better overall results.
Keywords: Ballistocardiogram, bed sensor, frequency analysis, heart rate monitoring, machine learning