If you wake up tired, it’s easy enough to deduce that you’ve had a bad night’s sleep. But measuring one’s sleep patterns in more detail usually requires polysomnography, which involves trying to fall asleep in a sleep clinic or hospital bed while wired up to sensors for heart rate, blood oxygen, body movement and other vital signs. Recording and understanding sleep patterns are vital for physical and mental health, especially for those suffering from sleep disorders, but polysomnography is cumbersome and expensive.
Now, researchers at A*STAR’s Institute of Infocomm Research (I2R), in collaboration with McLaren Applied Technologies, have combined wearable sensors and machine learning techniques to devise an alternative to polysomnography. Zhenghua Chen, lead researcher on the team, notes that movement sensors have become small and cheap enough that people routinely wear them throughout the day. As such, it should also be possible to use these sensors to record sleeping and waking through the night, then analyze the data for health insights using machine learning techniques.
“However, conventional machine learning approaches for sleep-wake detection require features of sleep-wake cycles to be manually defined by an expert, and some implicit features may be missed out,” Chen explained. To address this problem, the team developed a deep learning framework that automatically distinguishes sleep and wake phases based on accelerometer readings and heart rate variability (HRV) measurements derived from wearable devices.
The researchers first had to account for the high sampling rate of accelerometer readings, which creates many blocks of sequential time series data. Instead of analyzing all blocks at once, they applied a ‘divide and conquer’ strategy, whereby the dataset is broken into smaller segments and analyzed for local features representative of sleep and wakefulness. They then designed a ‘fusion framework’ to merge accelerometer data with HRV measurements to automatically define and detect sleep-wake cycles.
“Representative features of sleep-wake cycles can thus be automatically learned without human intervention. Compared with state-of-the-art methods, our approach improves the accuracy of sleep-wake detection by three to 19 percent. Hence, better detection performance can be achieved with our proposed method,” said Chen.
Going forward, Chen’s team plans to further develop their deep learning technique to identify more sleep stages, such as light sleep, deep sleep and rapid eye movement (REM) sleep. A finer understanding of sleep-wake cycles could help identify and diagnose many sleep disorders, said Chen.
The A*STAR-affiliated researchers contributing to this research are from the Institute of Infocomm Research (I2R).