[4L1-GS-10] AI application: Medicine / Healthcare

[4L1-GS-10-05] Development of technology to improve the accuracy of automatic sleep stage scoring by combining pre-training and clinical research data

〇yoshito tabata¹, Kotaro Ezure¹, Kentaro Morimoto², Yoshitake Kitanishi², Tomoko Yagi^3,4, Shintaro Chiba⁴, Shusuke Yoshimoto¹ (1. PGV Inc., 2. Shionogi & Co., Ltd., 3. Kurume University, 4. Ota Memorial Sleep Center)

Keywords:AI, Electroencephalography, Sleep, Healthcare

In recent years, sleep has been found to be associated with insomnia, hypersomnia and various diseases. Examining the state of sleep is pivotal in uncovering these issues, and a plethora of varied studies have been undertaken for this very purpose. Analysis of sleep stages using electroencephalography is a key metric extensively explored in research, which a crucial role in visualizing and scrutinizing the various states of sleep. Sleep stage scoring is dependent on the expertise and experience of professionals, presenting challenges such as a limited number of qualified scorers and time-intensive scoring processes. Progress has been developed in both research and products with a focus on leveraging deep learning for the automated scoring of sleep stages. In our previous research, we developed a sleep stage scoring system using a patch-type EEG sensor, achieving an accuracy of 78.6% and a Kappa coefficient of 0.70. We aimed to refine the sleep stage scoring inference engine established in the preliminary study. This paper shows two key achievements: (1) change the model architecture (2) training on a combination of our clinical research data and Open Datasets. The result shows a record accuracy of 83.1% and a Kappa coefficient of 0.749. This achievement represents a notable enhancement in accuracy compared to our previous study.