Turning over in bed, especially turning over at night, is a vital human unconscious behavior. Clinically, this movement disperses pressure between the body and bed, thus preventing bedsores. Several devices, such as acceleration and pressure sensors, can count turning overs automatically; however, they often require installation on the patients or in the bed. The simplest and noninvasive method to count turning overs is to record and count on video images, but this method cannot protect privacy. Images obtained using thermal sensors have been used to protect privacy; however, there are no reports of counting turning overs automatically using low-resolution sensors. We developed a novel device equipped with four low-resolution thermal sensors, with each sensor recording only an 8×8-pixel thermal image. The original data can protect patient privacy because the resolution is only ~28.8×28.8 cm per body, which is the lowest resolution compared to previous reports using thermal images. Using four sensors simultaneously enables us to collect sufficient data for automatic identification. We first used the bilinear interpolation method employed in a previous report to count turning overs; however, the results were unsatisfactory because turning overs produced extremely subtle changes in the original data compared with postural changes such as falls. After several attempts, we finally developed a unique identification program that interleaved all data from four sensors and then identified turning overs using residual neural network-18. Using the new system, the accuracy, recall, and precision of counting turning overs in bed improved to approximately 90% with an acceptable computation load in an experiment conducted on volunteers. This study demonstrated the feasibility of our device to count turning overs in clinical settings by the new identification program using four 8×8-pixel thermal images per frame, which have sufficiently low resolution to protect patient privacy.
Chou et al. (Thu,) studied this question.