[3B4-GS-11] AI and Society: medicine

[3B4-GS-11-02] A Preliminary Study of Machine Learning Algorithms for Predicting Depression Relapse from Life-Log Data

〇Ryo Kiguchi¹, Ayano Hata¹, Satoki Fujita¹, Yuki Yoshida¹, Yoshitake Kitanishi¹, Junichiro Yoshimoto², Aran Tajika³, Toshiaki A. Furukawa³ (1. Shionogi & Co., Ltd., 2. Nara Institute of Science and Technology, 3. Kyoto University)

[[Online]]

Keywords:Depression, Relapse prediction, BORUTA, Xgboost

The number of patients with depression has been increasing in recent years, and the decline in labor productivity caused by the characteristics of the disease is a major social problem. If relapse of depression can be detected by the change in the activity pattern preceding the relapse, we may be able to lead to early treatment. The purpose of this study was to explore algorithms suitable for the prediction of depression relapse by applying various machine learning algorithms (ex. BORUTA, Xgboost) to the data collected from the activity record application and the wearable device. The comparative results showed that some models could predict the severity level categorized based on K6 score (an index of psychological distress) with a considerable performance (kappa-coefficient >= 0.7; equivalent to ‘Substantial’ of Landis and Koch’s criteria), but that the performance could be extremely degraded to ‘Slight’ level of Landis and Koch’s criteria depending on machine learning algorithms. In this presentation, we also report the accuracy comparison by various machine learning approaches The study was conducted as the second analysis of the Fun to Learn, Act and Think through Technology-2 (FLATT2) study.