Sleep Posture Detection via Embedded Machine Learning on a Reduced Set of Pressure Sensors

• Published in: Sensors (Basel, Switzerland) Vol. 25; no. 2; p. 458
• Main Authors: Peruzzi, Giacomo, Galli, Alessandra, Giorgi, Giada, Pozzebon, Alessandro
• Format: Journal Article
• Language: English
• Published: Switzerland MDPI AG 14.01.2025; MDPI
• Subjects: Algorithms; artificial intelligence; Cameras; embedded machine learning; Embedded systems; Energy consumption; Humans; internet of things; Machine Learning; Neural networks; Posture - physiology; Pressure; pressure sensors; sensor selection; Sensors; Sleep - physiology; Sleep apnea; Sleep Apnea, Obstructive - physiopathology; support vector machine; Support vector machines; Textiles; Wearable computers; Wearable Electronic Devices; obstructive sleep apnea; sensor selection; support vector machine; internet of things; embedded machine learning; pressure sensors; artificial intelligence; sleep posture recognition
• ISSN: 1424-8220; 1424-8220
• DOI: 10.3390/s25020458

Sleep posture is a key factor in assessing sleep quality, especially for individuals with Obstructive Sleep Apnea (OSA), where the sleeping position directly affects breathing patterns: the side position alleviates symptoms, while the supine position exacerbates them. Accurate detection of sleep posture is essential in assessing and improving sleep quality. Automatic sleep posture detection systems, both wearable and non-wearable, have been developed to assess sleep quality. However, wearable solutions can be intrusive and affect sleep, while non-wearable systems, such as camera-based approaches and pressure sensor arrays, often face challenges related to privacy, cost, and computational complexity. The system in this paper proposes a microcontroller-based approach exploiting the execution of an embedded machine learning (ML) model for posture classification. By locally processing data from a minimal set of pressure sensors, the system avoids the need to transmit raw data to remote units, making it lightweight and suitable for real-time applications. Our results demonstrate that this approach maintains high classification accuracy (i.e., 0.90 and 0.96 for the configurations with 6 and 15 sensors, respectively) while reducing both hardware and computational requirements.