TL;DR
Researchers have developed SLeeP AIDePt-2, a computer vision model based on YOLOv5 that automatically detects sleeping positions in late pregnancy under real-world home conditions. By transitioning from controlled lab settings to messy, overnight home environments, the model achieves 82–89% accuracy in identifying the most common sleep positions, providing a critical tool for studying the link between supine (back) sleeping and adverse pregnancy outcomes like stillbirth.
What Problem Does This Paper Address?
Clinical evidence suggests that sleeping on one's back after 28 weeks of pregnancy is associated with late stillbirth and small-for-gestational-age infants. This is likely due to the heavy uterus compressing the mother’s major blood vessels (the inferior vena cava and aorta), which reduces blood flow to the fetus.
However, most existing data relies on self-reported sleep positions, which are notoriously unreliable due to recall bias. While wearable sensors (like accelerometers) exist, they can be uncomfortable, prone to displacement, and often fail to distinguish between the position of the thorax (chest) and the pelvis—a distinction that is physiologically vital in pregnancy. This paper addresses the need for a non-contact, objective, and automated way to monitor sleep positions in a natural home environment, even when participants are under thick blankets or sharing a bed.
Methods
The authors employed a deep learning approach using YOLOv5s (You Only Look Once, version 5 small), a state-of-the-art object detection framework.
Key technical choices included:
- Multi-Class/Multi-Participant Detection: The model was trained to detect 12 distinct sleeping positions, a sitting position, and pillows for both the pregnant person and their partner simultaneously.
- Disabling Flip Augmentation: Standard CV training often "flips" images to increase data variety. The authors disabled this because, in pregnancy, the distinction between left and right is anatomically critical (the aorta and vena cava are not centrally located).
- Weighted Random Sampler: To combat "class imbalance" (since people spend much more time in common positions like "left lateral" than rare "hybrid" tilts), they used a sampler that forced the model to "see" rare positions more frequently during training.
- Infrared Processing: Since most sleep occurs in the dark, the model was trained primarily on infrared (grayscale) video frames.
Datasets
The study utilized a combined dataset of 12,930 images with 47,001 manual annotations:
- Controlled-Setting Dataset: 24 participants simulating specific positions in a lab-like home environment (from the previous "SLeeP AIDePt-1" study).
- Real-World Dataset: Overnight recordings from 15 pregnant participants and 13 bed partners in their own homes. This included "messy" data: thick duvets, patterned sheets, pets, and children entering the bed.
Key Insights and Ideas
- Pelvis vs. Thorax: Unlike most sleep trackers that treat the body as a single block, this model attempts to classify positions where the chest might be flat (supine) but the pelvis is tilted. This is a "pregnancy-first" design choice, as pelvic orientation is the primary driver of uterine vessel compression.
- The "Messy" Reality of Home: The transition from version 1 to version 2 of this model highlighted how much "noise" exists in real homes. The model had to learn to "see" through thick duvets and ignore patterned bedsheets that often disappear or change appearance under infrared light.
- Partner Presence: A major hurdle for computer vision in the bedroom is the bed partner. SLeeP AIDePt-2 successfully detects both individuals simultaneously, ensuring the partner's movements aren't misidentified as the pregnant participant's.
Key Results
The model showed significant improvement over its predecessor when tested against real-world complexity:
- Common Positions: For the three most frequent positions (left lateral, right lateral, and supine), the model achieved an Average Precision ([email protected]) of 0.82 to 0.89.
- Sitting: The model was exceptionally good at detecting sitting up at the edge of the bed (AP of 0.92).
- Overall Performance: Across all 14 classes (including rare/hybrid tilts), the mean Average Precision ([email protected]) was 0.70.
- Recall: The model had a 15-to-19% chance of failing to detect a common position when present, a relatively low false-negative rate for such a complex environment.
Limitations and Future Work
The authors identify several areas for improvement:
- Camera Placement: The model is sensitive to the camera's angle. If the camera isn't centered at the head of the bed at a specific height (1.6–1.7m), performance drops.
- Rare Positions: Performance was significantly lower for "hybrid" positions (e.g., supine thorax with pelvic tilt), likely due to a lack of sufficient real-world training examples.
- Pets: While pets were annotated, there wasn't enough data for the model to reliably distinguish a dog under the covers from a human limb.
Next Steps: The researchers are launching The DOSAGE Study, an international prospective cohort study. They plan to use this model to quantify the "dose" of supine sleep (how many minutes/hours) and its actual impact on fetal growth and stillbirth risk, moving beyond simple "going-to-sleep" surveys.
Summary updated