Non-Intrusive Sensor-Based Sleep Monitoring and Machine Learning Classification for Individuals with Disabilities in Residential Facilities: A Case Study from the MetaSalute Project

The growing ageing population calls for non-intrusive monitoring frameworks capable of supporting the continuous assessment of frail individuals in real-world care settings. This paper presents a machine-learning-based approach for classifying nocturnal health conditions using physiological measurements acquired through bed-based sensors in long-term care facilities. Developed within the MetaSalute project, the study considers data collected from November 2025 to March 2026 through bed-bands installed in three residential care facilities, allowing the monitoring of 27 subjects. Caregiver-reported questionnaires were used to derive binary labels by assigning a class to each answer. Heart rate and respiratory rate signals were aggregated into one-hour windows, transformed into statistical descriptors, and normalized on a subject-specific basis. A Balanced Random Forest classifier was then trained to address the class imbalance typical of real-world clinical monitoring data. The model achieved a balanced accuracy of 79.7% and a ROC AUC of 0.84, showing good capability in detecting critical nocturnal events.