Sleep Position Classification using Transfer Learning for Bed-based Pressure Sensors
Olivier Papillon, Rafik Goubran, Life Fellow IEEE, James Green, Senior Member IEEE, Julien Larivière-Chartier, Caitlin Higginson, Frank Knoefel, MD, Rébecca Robillard · 2025
Read Sleep Position Classification using Transfer Learning for Bed-based Pressure Sensors by Olivier Papillon, Rafik Goubran, Life Fellow IEEE, James Green, Senior Member IEEE, Julien Larivière-Chartier, Caitlin Higginson, Frank Knoefel, MD, Rébecca Robillard (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-08111.txt.
Introduction
Papillon and co-authors classify sleep positions from a pressure mat under a mattress, using transfer learning. The mat tracks patients without wires or wearables. Sleep posture affects sleep quality and disorders like apnea. Data came from a clinic's low-resolution mat in a sleep lab. Patients had suspected sleep disorders. Labels for training such models are scarce in clinics.
The team adapted vision models pre-trained on images with a masked learning method. Transfer let the models learn from the small clinic set. Four sleep positions were the target classes. Borrowed features beat training from scratch in this low-data setting. The approach suits hospitals that cannot gather huge labeled datasets.
Catalog record
| CREATOR | Olivier Papillon, Rafik Goubran, Life Fellow IEEE, James Green, Senior Member IEEE, Julien Larivière-Chartier, Caitlin Higginson, Frank Knoefel, MD, Rébecca Robillard |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 25 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-08111
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Sleep Position Classification using Transfer Learning for Bed-based Pressure Sensors Olivier Papillon, Rafik Goubran, Life Fellow IEEE, James Green, Senior Member IEEE, Julien Larivière-Chartier, Caitlin Higginson, Frank Knoefel, MD, Rébecca Robillard Abstract Bed-based pressure-sensitive mats (PSMs) offer a non-intrusive way of monitoring patients during sleep. We focus on four-way sleep position classification using data collected from a PSM placed under a mattress in a sleep clinic. Sleep positions can affect sleep quality and the prevalence of sleep disorders, such as apnea. Measurements were performed on patients with suspected sleep disorders referred for assessments at a sleep clinic. Training deep learning models can be challenging in clinical settings due to the need for large amounts of labeled data. To overcome the shortage of labeled training data, we utilize transfer learning to adapt pre-trained deep learning models to accurately estimate sleep positions from a low-resolution PSM dataset collected in a polysomnography sleep lab. Our approach leverages Vision Transformer models pre-trained on ImageNet using masked autoencoding (ViTMAE) and a pre-trained model for human pose estimation (ViTPose). These approaches outperform previous work from PSM-based sleep pose classification using deep learning (TCN) as well as traditional machine learning models (SVM, XGBoost, Random Forest) that use engineered features. We evaluate the performance of sleep position classification from 112 nights of patient recordings and validate it on a higher resolution 13-patient dataset. Despite the challenges of differentiating between sleep positions from low-resolution PSM data, our approach shows promise for real-world deployment in clinical settings. I Introduction With age comes a greater reliance on healthcare services as age-related health conditions become more prevalent. Remote health monitoring could reduce the strain on the Canadian healthcare system. [ 1 ] Polysomnography, or the study of sleep, becomes increasingly important as the prevalence of sleep-related disorders increases with age. Researchers have proposed using bed-based pressure-sensitive mats (PSM) placed under the mattress as an unobtrusive method of monitoring patients. To this end, bed-based PSM have been installed at the Royal Ottawa Hospital sleep study lab, where they record time-varying contact pressure data throughout the night. This technology has proven useful for monitoring bed-based mobility [ 2 ] , sleep stages [ 3 , 4 , 5 , 6 , 7 , 8 ] , pose estimation [ 5 , 6 , 10 , 11 , 7 , 9 ] , and respiration rate [ 12 , 13 ] . In the present study, we focus on automatically recognizing body positions from data recorded by the PSM. Research has indicated that an individual’s sleep position significantly influences their sleep quality, and the prevalence and severity of sleep disorders such as sleep apnea [ 14 ] . As such, we are interested in accurately and automatically recognizing the body positions of sleeping patients using only PSM data. Accurate pose estimation enables researchers to select the most appropriate estimation approaches for downstream tasks, such as respiration rate estimation or patient movement detection. For example, robustly estimating respiration rate from PSM data requires different approaches for side-lying subjects vs. supine; by providing the ability to determine a patient’s pose reliably, we improve downstream respiration rate estimation and achieve clinical benefit. To do this, we use machine learning (ML) to analyze PSM data and extract meaningful patterns and valuable insights about the data. Recent research has extracted sleep positions from PSM data using ML algorithms such as Support Vector Machines (SVM) and various Deep Learning (DL) architectures [ 3 ] , [ 4 ] - [ 6 ] . These methods are effective; however, recent DL architectures have demonstrated improved classification and computer vision task performance. Since PSM data represent images over time (i.e., video), the use of DL is expected to outperform traditional ML that requires hand-engineered feature extraction. Although DL methods can be very powerful, training DL models still poses a significant challenge due to the large amount of labeled data required for training [ 15 ] . To overcome the present shortage of labeled training data, we can utilize transfer learning to leverage state-of-the-art pre-trained DL models and fine-tune them for our specific task. This research makes contributions to the fields of computer vision, machine learning, and medical measurements and instrumentation by examining the effectiveness of newer DL models to detect sleeping poses from PSM data. Given that PSMs represent a privacy-preserving sensor modality, such a system could ultimately be deployed in a real-time sleep lab or for remote monitoring at home and long-term care facilities. II Related Works II-A Posture Recognition using Bed-Based Pressure Sensor Arrays In [ 6 ] , Foubert et al. performed sleep position classification using classical ML methods to classify the sleep positions. The study used bed-based PSM with 44 sensors and 25 participants of various ages. The researchers extracted features from the PSM data (e.g., the sum of sensor values, the center of pressure) to train SVM and K-NN classifiers to recognize six different sleeping poses: Prone, Supine, Left Lateral, Right Lateral, Left Fetal, and Right Fetal. In the present study, this method served as a baseline for comparison of the effectiveness of the newer pre-trained DL architectures. II-B Sleep-Wake and Body Position Classification with Deep Learning using Pressure Sensor Mat Measurements In [ 5 ] , Green et al. performed sleep-wake and body position estimation using a TCN (Temporal Convolution Network) with PSM data similar to those used in the present study. PSM data and gold standard body positions were annotated at the Royal Ottawa Hospital. A low-resolution PSM was used that comprised only 72 sensors. Only three body positions were included in the dataset since the Supine and Prone positions were merged into a single position to simplify the training of the classifiers. The present study expands on this study by doubling the number of sensors (144 sensors) and treated the problem as a 4-way classification rather than 3-way. By including Prone as a distinct body position, the problem becomes substantially more difficult since this position represents a minority class and is difficult to distinguish from other poses based on PSM data alone. II-C A Pressure Map Dataset for In-bed Posture Classification In [ 11 ] , Pouyan et al. presented the publicly available Pmat dataset. This dataset was collected from 13 participants laying on a Force Sensitive Application pressure mapping mattress comprising a 32 × 64 grid of sensels (2048 total) with approximately 1-inch separation. The dataset included pressure maps for adults in 17 different postural poses, which were categorized into three classes: Supine, Right, and Left. This dataset has been used by a variety of research groups for pose estimation, sleeper identification, and other tasks. Of note, Doan et al. [ 16 ] and Davoodnia and Etemad [ 17 ] presented leave-one-subject-out results for 3-way sleep position estimation from PSM exceeding 95% accuracy. In [ 17 ] , Davodnia and Etemad present simultaneous identity and posture recognition in smart beds using deep multitask learning techniques. In [ 16 ] , Doan et al. used an EfficentNet B0 based classifier to identify 17 sleeping postures as patients rested in the hospital beds. The research highlighted the potential of real-time posture recognition systems in enhancing patient care and optimizing bed utilization in hospital settings. In the