Smart Furniture-Embedded Bed-Interaction and Multimodal Room Sensing with Sleep-Aware Hierarchical Fusion for Control-Oriented Bedroom State Recognition

Abstract

Highlights What are the main findings? A bed-centered multimodal sensing framework was developed by integrating furniture-embedded bed-interaction sensing with bedside IEQ, room-context, and time features to recognize eight control-oriented sleep-related bedroom states. The proposed SA-HMF model achieved strong subject-independent performance, with 0.925 Accuracy, 0.874 Macro-F1, and reliable extraction of in-bed, final out-of-bed, nocturnal out-of-bed, and wake-up transition events. What are the implications of the main findings? Furniture-embedded bed-interaction sensing provides essential in-bed and movement information, while room-context and time cues remain necessary for distinguishing unoccupied and occupied-out-of-bed states. The edge-deployable SA-HMF model supports privacy-preserving local inference and exploratory downstream smart bedroom applications. Abstract Reliable control-oriented bedroom state recognition is needed because room-level occupancy sensing cannot distinguish in-bed, out-of-bed, movement, nocturnal bed-exit, and wake-related conditions. This study developed a privacy-preserving bed-centered sensing framework and a sleep-aware hierarchical multimodal fusion model, termed SA-HMF, for recognizing eight operational bedroom states (S0–S7). SA-HMF uses a bed-conditioned, quality-aware asymmetric architecture: bed-load and bed-mounted accelerometry are first encoded into a bed-interaction representation, which then conditions the fusion of bedside indoor environmental quality, room-context, temporal, and sensing quality information. Data were collected from 20 healthy adults across 127 valid participant-nights in one experimental bedroom. The recordings generated 342,684 overlapping 30 s windows using a 10 s stride; these windows were treated as correlated model instances rather than independent observations. Under the primary subject-independent split, SA-HMF achieved 0.925 Accuracy, 0.874 Macro-F1, 0.927 Weighted-F1, and 0.864 Balanced Accuracy. LOSO evaluation yielded a Macro-F1 of 0.852 ± 0.043, with a participant-level 95% confidence interval of 0.833–0.870. Using non-overlapping 30 s windows, SA-HMF retained 0.918 Accuracy, 0.865 Macro-F1, and 0.856 Balanced Accuracy. Retrospective event extraction F1-scores ranged from 0.846 to 0.950. The optimized model was 2.7 MB and required 15.6 ± 3.2 ms per inference. These results support feasibility and transfer to unseen healthy participants within the tested bedroom and sensing configuration. External validation across bedrooms, beds, sensing installations, and populations remains necessary.