Sensor-Stack Limits on Contactless In-Bed Body Position: A 20-Subject Multimodal Radar + Thermal LOSO Characterization

Abstract

Contactless in-bed body-position inference can be limited by exposed sensor representation rather than classifier choice. We characterize a bedside 60 GHz frequency-modulated continuous-wave (FMCW) radar with on-device constant-false-alarm-rate (CFAR) point-cloud output plus a low-resolution (24 x 32 nominal, rows x columns) thermal array on two leave-one-subject-out (LOSO) evaluations derived from the same 20-subject cohort: 273 supervised in-bed posture holds (148.7 minutes) and an enter/exit bed-presence audit from the same cohort. The cohort is a 20-subject friends-and-family calibration sample (13 minors, ages 5-68; 8 residences), so these are characterization figures on a convenience cohort, not population-level performance. The motivating use case is prone-position monitoring, because prone position has been associated with sudden unexpected death in epilepsy (SUDEP) in retrospective studies. Fused radar + thermal logistic regression reaches a 0.871 median leave-one-subject-out balanced accuracy for in-bed vs out-of-bed classification. For four-class posture, the best tested pipeline (a full-feature stacked ensemble) reaches 0.674 aggregate balanced accuracy. Prone recall is 0.50 and prone precision is 0.41, so this is not deployable prone detection. Ablations show that thermal solves left-vs-right discrimination (radar-only lateral swaps 35-42%; thermal-only ~8%), but the expected supine-vs-prone breathing cue appears only as a class-level aggregate shift in CFAR output (Cohen's d=0.61), with clean per-hold peaks in 8.4% of holds. The thermal array's usable resolution in this cache was half its nominal column count, too coarse to separate face-from-back-of-head signatures. The results point to raw range-FFT access, rather than classifier tuning on CFAR detections, as the next hardware experiment.