A study led by the University of Western Australia (UWA) indicates machine learning-based algorithms they developed for 3-D facial photography could provide a simple and highly accurate method of predicting the presence of sleep apnoea.

The study, published in the Journal of Clinical Sleep Medicine, suggested that it might also be possible to predict the severity of a person’s sleep apnoea from these photographs.

Feeding machine learning

The complete system is based on machine learning technology, which means researchers feed a machine learning algorithm with features – 3-D distances between certain important landmarks on the face – of non-sleep apnoea (control) and sleep apnoea patients, and then train it to learn the difference between the two classes.

Once trained, the machine learning algorithm was able to distinguish between the controls and patients with an accuracy of more than 91%.

Dr. Syed Zulqarnain Gilani, from UWA’s School of Computer Science and Software Engineering, told HealthcareITNews the biggest challenge in solving any such problem is collection of relevant data.

“That includes the 3-D images of patients, their sleep pattern and behaviour as well as a myriad of other variables,” he explained. “Our team started collecting these data in early 2016. We also required the 3-D face data of non-sleep apnoea individuals from the general population.”

Sleep disorders cost billions

Dr. Gilani noted sleep disorders are estimated to cost the Australian health system more than $5 billion annually, and said more than half the cost is associated with sleep apnoea, which is associated with snoring and repeated periods of ‘choking’ during sleep.

Sleep apnoea also causes daytime sleepiness and is strongly linked to sleepiness related accidents, diabetes, cardiovascular diseases and depression.

“Despite sleep apnoea being treatable, the vast majority – up to75% – of individuals remain undiagnosed,” he said. “This is largely because current methods of assessing sleep apnoea are expensive and access to them is limited. The development of this technology will significantly reduce the costs of possible early detection of sleep apnoea.”

Dr. Gilani pointed out the team’s work builds on previous studies that helped identify that the structure of the face, head and neck, which have already played a key role in diagnosing sleep apnoea.

Perfecting 3-D mapping

He said the research teams aim to perfect the 3-D mapping technology using more data and robust algorithms to make it a reliable tool for diagnosing obstructive sleep apnoea.

“Although our current study uses only 400 3-D faces, we have already collected more than 2,000 faces of sleep apnoea patients,” he said. “We want to develop a technology that is cheap, easily accessible and reliable.”



White papers