Transforming patient care: AI-powered sleep posture classification for pressure injury prevention

Background

Pressure injuries (bedsores) remain a significant and costly healthcare concern, particularly for bedridden or mobility-impaired patients. Early detection and continuous monitoring of sleep posture are essential for effective prevention; however, existing systems are often expensive, intrusive, or lack sufficient accuracy.

Research question

This study investigates whether a wearable IMU sensor-based system integrated with a lightweight deep learning model—SleepPosNet—can accurately classify five common sleeping postures and outperform traditional learning models.

Methods & results

Data from 100 participants (18–65 years; 16 male/84 female) were collected using three IMU sensors (chest, right leg, left leg). Tri-axial accelerometer, gyroscope, and magnetometer data were fused into nine Euler-angle channels and labeled into five posture classes. A lightweight 1D-CNN (SleepPosNet) was trained (Adam, lr = 1e-3, batch = 64, 30 epochs) and evaluated with stratified 70–30, 80–20, and 90–10 splits, achieving up to 98.94 % accuracy, consistently surpassing MLP, Naïve Bayes, and Logistic Regression. In a 10-fold cross-validation with deep learning baselines (BiLSTM, LSTM, GRU), SleepPosNet reached 97.39 % accuracy with only ∼ 13 k parameters, the shortest epoch time (∼28.6 s), low latency (∼0.239 ms/sample), and high throughput (∼4.19 k samples/s). While BiLSTM achieved slightly higher accuracy (98.34 %), it required far greater computation. SleepPosNet thus offers the best accuracy–efficiency trade-off for embedded and real-time applications.

Significance

SleepPosNet offers a non-invasive, low-cost, and highly accurate solution for real-time sleep posture monitoring. Its lightweight structure makes it suitable for deployment in hospital and home care settings, with the potential to reduce healthcare costs and improve outcomes by aiding in the prevention of pressure injuries.