HyPoNet: Fine-Grained Sleep Posture Recognition from a Single Abdominal Accelerometer

Abstract

Fine-grained sleep posture recognition is essential for the non-invasive management of conditions such as gastroesophageal reflux disease (GERD) and obstructive sleep apnea (OSA). Traditional systems typically recognize only a limited set of coarse sleep positions, thereby restricting their clinical effectiveness in real-world scenarios. This study presents HyPoNet, a lightweight deep learning model designed to classify twelve distinct sleep postures using data from a single wearable sensor system. The proposed hardware platform consists of a tri-axial accelerometer (ADXL345) positioned on the abdomen, interfaced with a low-power microcontroller unit (ESP32) for real-time signal acquisition and wireless data transmission. Acceleration signals along the x, y, and z axes were collected from ten healthy participants performing twelve predefined sleep positions under controlled conditions. The collected data were segmented using a sliding window method, and a subject-independent evaluation strategy was applied: data from eight participants were used for training and validation (in an 80:20 split), while data from the remaining two participants were reserved for testing. HyPoNet employs a hybrid neural network architecture combining one-dimensional convolutional layers for spatial feature extraction with bidirectional long short-term memory (BiLSTM) units to model temporal dependencies in the acceleration signals. The model achieved a mean accuracy of 97.29% and an average F1-score of 90.72%, outperforming baseline models including CNN, GRU, and Transformer-based approaches. With its low computational footprint and high classification performance, HyPoNet offers a promising solution for embedded sleep posture monitoring in home-based and clinical settings.