Sina Akbarian 1,2, Ghazaleh Delfi 1,2, Kaiyin Zhu 1, Azadeh Yadollahi 1,2, Babak Taati 1,2,3,4
1 Toronto Rehabilitation Institute, University Health Network
2 Institute of Biomaterials and Biomedical Engineering, University of Toronto
3 Department of Computer Science, University of Toronto
4 Vector Institute, Toronto, Ontario, Canada
Obstructive sleep apnea (OSA) is a respiratory disorder characterized by interruption to breathing during sleep. It is estimated that about 10% of the population suffer from OSA. OSA is a risk factor for cardiovascular diseases, stroke, and abnormal glucose metabolism. A recent study found that the overall healthcare costs of untreated OSA patients were nearly 25% higher than those patients undergoing treatment. The severity of OSA is often associated with sleeping in the supine body position. In addition, changing the head position to lateral also plays an important role in decreasing the OSA. Positional therapy could have an impact on decreasing the occurrence of OSA. Positional therapy involves wearing items, such as backpack, to encourage sleeping in the lateral position. Although positional therapy is simple, its limitation is that it is difficult or even impossible to know for certain whether the patient remained in the desired position throughout the night. Therefore, monitoring the head/body position could provide feedback to patients and physicians. In addition, analyzing changes in sleep position could have applications in assessing sleep quality.
In this work, three approaches were used to detect head position during sleep. In the first two, supervised classifiers were trained to estimate head position based on features extracted from infrared images. In the third method, three different convolutional neural network structures were trained to predict head position during sleep. To detect body position, the same convolutional neural network architectures were trained. The models were trained on overnight sleeping data of 15 participants with OSA and tested on overnight sleeping data of 19 participants with OSA, collected in a different laboratory room. The best performing (DarkNet19) correctly estimated lateral vs. supine head position with 92% accuracy and 94% F1-Score, and the lateral vs. supine body position with 87% accuracy, and 87% F1-Score.