Non-contact Detection of Head Posture During Sleep

Sina Akbarian (1,2), Kaiyin Zhu (1), Nasim Montazeri (1),

Azadeh Yadollahi (1,2,3), Babak Taati (1,2,3,4)

1. Toronto Rehabilitation Institute

2. Institute of Biomaterials and Biomedical Engineering, University of Toronto

3. Department of Computer Science, University of Toronto

4. Vector Institute, Toronto, Ontario, Canada

Introduction: Obstructive sleep apnea (OSA) is a respiratory disorder characterized by the collapse of pharyngeal airway during sleep. About 10% of the population have OSA. The severity of OSA is often associated with sleeping in the supine position. Therefore, positional therapy that involves wearing items, such as backpack, to encourage sleeping in the lateral position, is a simple treatment for OSA. The current studies monitor the effects of supine posture of whole body (trunk) on sleep apnea severity. However, recent studies have shown that head position has more effect on sleep apnea severity than the trunk position. The available approach to monitor the head position during sleep is to attach an accelerometer sensor to the head. However, attaching sensor to head is inconvenient during sleep. Therefore, the goal of this study was to use a non-contact algorithm based on processing of infrared video to detect head position during sleep.
Methods: Individuals referred for diagnosis of sleep apnea in the sleep laboratory of Toronto Rehabilitations Institute were recruited. Simultaneously with full polysomnography, infrared video of overnight sleep was recorded. The gold standard of the head position is manual annotation. Three machine learning approaches were used to detect head position during sleep. In the first method (PF), Haar feature-based cascade classifiers were used to detect nine binary features. The second method (LF) used the 3D coordinates of 68 facial landmarks detected via the hourglass deep convolutional network. Landmark points were normalized to the nose length. Extracted features from each method were used to train a binary random forest classifier to detect supine vs. lateral head positions. In the third method, a convolutional neural network (DarkNet19) was trained on the whole recorded images to predict head position during sleep.
Result: Fifty subjects, age: 53 ± 15 years, BMI: 29 ± 6 kg/m2, 30 men and 20 women, AHI: 25 ± 29 event/hour, sleep duration: 5 ± 1 hours, were included in this study. The PF and LF were respectively estimated head position with 59.9% and 66.3% accuracy and 72.5% and 77.7% F1-Score. Both methods had poor performance in detecting the head position, as compared to the Darknet19 estimated head position with 88.3% accuracy and 91.3% F1-Score.
Conclusion: In this work, we developed a highly accurate algorithm that can automatically detect head position during sleep and provide feedback for positional therapy.