Accurate Sleep Stage Measurement at Home... Development of an Artificial Intelligence Model

Bundang Seoul National University Hospital Professor Yoon In-Young's Team
Incorporates Home Breath Sound Data
10%P Accuracy Improvement Over Previous Models

The research team led by Professor Yoon In-young from the Department of Psychiatry at Bundang Seoul National University Hospital (co-corresponding author Dr. Kim Dae-woo, head of AI at A-Sleep) announced on the 7th that they have developed an artificial intelligence (AI) model capable of measuring sleep stages with high accuracy at home.

Professor Inyoung Yoon (left), Department of Psychiatry, Seoul National University Bundang Hospital, and Dr. Daewoo Kim, AI Director at A-Sleep.

During sleep, a person goes through sleep stages including wakefulness → light sleep → REM (Rapid Eye Movement) sleep → deep sleep. In normal sleep, each stage is observed in a certain proportion, serving functions such as relieving fatigue and storing memories. However, if the normal structure of sleep changes due to physical or psychological factors, the quality of sleep deteriorates, and in severe cases, it can lead to sleep-related disorders, requiring caution.

The research team trained the AI model on various sounds occurring during sleep at home by utilizing 6,600 hours of sound data recorded via smartphones during sleep, home polysomnography data, and 270 hours of breathing sound data recorded through smartphones during home polysomnography.

Previously developed models were based on polysomnography results conducted in hospitals and had limitations in properly reflecting noises and events occurring at home. By using home polysomnography results, which are closer to actual sleep environments, the accuracy was significantly improved. The newly developed AI model demonstrated about a 10 percentage point higher performance compared to existing models.

Professor Yoon said, "We have proven that sleep stages can be measured with high accuracy in home environments compared to existing AI models based on hospital settings," adding, "By using the newly developed model to understand usual sleep patterns, it is expected to help in early diagnosis of patients who may develop sleep-related disorders and assist them in receiving active treatment."

The research findings were introduced at the 'SLEEP 2023' academic conference hosted by the American Academy of Sleep Medicine starting on the 3rd, and at the AI conference 'CLR.' They were also published in the latest issue of 'JMIR (Journal of Medical Internet Research),' a prestigious international journal in the field of health informatics.