TL;DR
This paper introduces a non-obtrusive sleep posture recognition system using a triple Ultra-Wideband (UWB) radar setup combined with Vision Transformers (ViTs). The system aims to accurately identify sleep postures, even through blankets, to aid in the assessment and management of Obstructive Sleep Apnea (OSA). The best performing configuration, a Swin Transformer with side and head radar, achieved an impressive 80.8% accuracy.
What Problem Does This Paper Address?
Obstructive Sleep Apnea (OSA) is a common sleep disorder with significant health consequences. Sleep posture is a known factor influencing OSA severity, with supine positions often exacerbating the condition. Therefore, accurately monitoring sleep posture is crucial for OSA assessment and rehabilitation. Existing methods have drawbacks: contact-based sensors can disrupt sleep, and camera-based systems raise privacy concerns. While radar systems offer a promising non-contact, privacy-preserving alternative, challenges remain in accurately distinguishing subtle posture changes, especially when individuals are covered by blankets. This paper tackles these limitations by proposing a multi-radar UWB system and leveraging advanced deep learning models to achieve robust blanket-penetrating sleep posture recognition.
Methods
The core of this research lies in the innovative combination of hardware and software.
- Hardware: A triple Ultra-Wideband (UWB) radar system was employed. Three UWB radar sensors were strategically positioned: one above the bed (top), one at the side, and one near the head of the participant. This multi-sensor approach was designed to capture richer spatial and temporal information about the body's shape and movements.
- Data Processing: Raw radar signals were preprocessed to remove static objects and environmental noise. Crucially, the data from each radar was transformed into an "imitated image" format. For single-radar configurations, the data was replicated across RGB channels. For dual and tri-radar setups, data from different radars were assigned to different RGB channels, effectively creating a multi-channel image representation. Data augmentation techniques like scaling, magnitude warping, and time warping were applied to increase the dataset's diversity.
- Machine Learning Models: The processed radar data, now in image format, was fed into various deep learning models. The study evaluated both Convolutional Neural Networks (CNNs) like ResNet50, DenseNet121, and EfficientNetV2, and Vision Transformer (ViT) models, specifically the traditional ViT and the Swin Transformer V2. ViTs, known for their ability to capture global representations through self-attention mechanisms, were hypothesized to be particularly effective for this task.
Datasets
The study utilized data collected from 30 healthy adult participants (19 males, 11 females) aged 18-27. Participants performed four recumbent postures: supine, left side-lying, right side-lying, and prone. Each posture was recorded for 15 seconds, and this process was repeated ten times for each participant. This resulted in a total of 1200 samples. The dataset was split for training (18 participants), validation (6 participants), and testing (6 participants). The paper does not introduce a new public benchmark dataset but rather uses the collected data for evaluation.
Key Insights and Ideas
- Multi-Radar Synergy: Combining data from multiple radar sensors (especially side and head) significantly improved posture recognition accuracy compared to single-radar setups. This suggests that different vantage points provide complementary information crucial for distinguishing subtle posture variations.
- Vision Transformers Excel: The study found that transformer-based models, particularly the Swin Transformer, outperformed CNN-based models. This highlights the ViT's strength in capturing global contextual information, which is vital for understanding the overall body shape and posture, even with the lower-resolution "imitated images" derived from radar data.
- Blanket Penetration is Achievable: The successful recognition of postures through blankets demonstrates the viability of UWB radar for unobtrusive sleep monitoring, addressing a key limitation of visual or contact-based methods.
- Top Radar Limitations: The top radar configuration consistently yielded lower accuracy. This is attributed to its distant placement, leading to higher signal attenuation and a poorer signal-to-noise ratio, which negatively impacts prediction.
Key Results
The research evaluated various radar configurations and deep learning models.
- Best Performing Model: The Swin Transformer V2 model, when used with the side + head radar configuration, achieved the highest prediction accuracy of 0.808 on the test dataset.
- Model Performance: Transformer-based models (average accuracy 0.613-0.637) generally outperformed CNN-based models (average accuracy 0.538-0.551).
- Radar Configuration Performance: Dual-radar configurations generally performed better than single-radar configurations. The side + head configuration was the most effective among the dual-radar setups. The tri-radar configuration (top + side + head) showed promise but did not surpass the best dual-radar configuration.
- Posture Accuracy: The confusion matrix revealed that side-lying postures were recognized with very high accuracy (110/120 correct predictions), while supine postures also performed well. Prone postures showed the lowest accuracy among the four.
Limitations and Future Work
The authors acknowledge several limitations:
- Limited Dataset Size: While data augmentation was used, a larger dataset with more diverse participants (varying ages, sexes, body builds) would enhance the model's robustness and generalizability.
- Incomplete Spatial Information: The current approach collapses radar signals into a 1D time-series, losing detailed spatial information about body topology.
- Vital Sign Interference: The system did not explicitly isolate vital sign signals (like respiration), and it's unclear if the ViT used these as salient features.
Future work suggested includes:
- Synthetic Aperture Radar (SAR): Exploring SAR techniques to capture more complete spatial information.
- Larger Datasets: Expanding the dataset with more participants and variations.
- Advanced Modeling: Investigating techniques like DensePose for more detailed human pose estimation.
- Investigating Vital Signs: Understanding the role of vital signs in posture estimation and potentially isolating them.
Summary updated