Ballistocardiography (BCG) enables unobtrusive cardiac monitoring using bed-integrated sensors but remains highly sensitive to body posture, leading to variability in waveform morphology and degraded downstream analysis. In real-world home monitoring, posture information is typically unavailable. In this study, we investigate whether body posture can be automatically detected directly from BCG signals using machine learning.
Fifteen healthy participants were recorded using an Emfit bed sensor in five standardized postures: supine, prone, left lateral, right lateral, and Semi-Fowler position. Each posture was measured for approximately five minutes and repeated three times. Signals were synchronized with reference electrocardiography and manually annotated for signal quality. BCG recordings were segmented into fixed-length windows and characterized using morphological, time- and frequency-domain features.
Supervised classification models were trained to identify posture directly from BCG signals using subject-independent validation to evaluate generalization across individuals. Feature attribution analysis was performed to investigate which waveform characteristics contributed most strongly to posture detection.
Results demonstrate that posture information can be reliably extracted from BCG recordings, indicating that mechanical cardiac signals contain sufficient information for posture inference. The proposed approach enables posture-aware cardiac monitoring and may improve robustness of heart rate and heart rate variability estimation in long-term unobtrusive sleep monitoring environments.