An artificial intelligence model can diagnose obstructive sleep apnoea (OSA) with nearly 90 per cent accuracy using only a night-vision camera, potentially doing away with the need for intrusive tests in specialist laboratories.
OSA occurs when the walls of the throat relax and narrow during sleep, which has been linked to high blood pressure, type 2 diabetes and heart disease. The standard method of diagnosis is polysomnography (PSG), which involves an overnight hospital stay with as many as 20 sensors attached to various body parts.
Information from these sensors is then compressed down into a single metric called the apnoea-hypopnea index (AHI), based on the number of apnoeas 鈥� periods where a person stops breathing 鈥� per hour. This is then used to determine the severity of the condition.
Advertisement
The uncomfortable and unfamiliar process of PSG can make people act differently than normal, affecting diagnoses. A one-night observation can also be inaccurate because OSA鈥檚 severity can vary from night to night.
鈥淢any people say that they cannot sleep well during PSG due to the inconvenience,鈥� says at Seoul National University in South Korea. 鈥淎ccurate testing requires multi-night observation of natural sleep to mitigate night-to-night sleep variability and first-night effect, which is nearly impossible for PSG in practice.鈥�
Now, Kim and his colleagues have developed a technique where a person suspected of having OSA is monitored by an infrared camera while they sleep. Their AI tool, called SlAction, can then diagnose the condition from the video alone. Crucially, such tests can be done cheaply, over many nights and potentially in the person鈥檚 home.
The researchers used video clips of hundreds of people 鈥撎齛bout 6 hours long per person 鈥� that were collected in three hospitals. These were annotated with the actual diagnosis from human experts in order to train the AI system to recognise visual clues of OSA, which may include gasping and waking a lot in the night. When testing the model on similar clips that it hadn鈥檛 seen before, they found the system could diagnose OSA with 88 per cent accuracy.
at Flinders University in Australia says approaches such as using infrared video are promising, given that PSG gets OSA鈥檚 severity level wrong between 20 per cent and 50 per cent of the time due to night-to-night variability. 鈥淚t鈥檚 a time-consuming, expensive, inaccurate, labour-intensive process,鈥� he says.
Eckert believes that cheaper, simpler methods to diagnose sleep apnoea in peoples鈥� homes, where tests can be done over many nights, are the way forward. He adds, however, that SlAction would still use the simplistic metric of AHI to gauge OSA鈥檚 severity, which needs updating.听
arXiv
