Presentation Information
[P-1-06]Development of Machine Learning-Based Computer-Aided Diagnosis System for Psychiatric Disorders Using Wearable Physiological Data
*Goomin Kwon1,2, Yuhwa Hong1, Hosang Moon1, Miseon Shim1, Seunghwan Lee2,3,4 (1.Tech Univ of Korea(Korea), 2.CEClab(Korea), 3.Bwave(Korea), 4.Ilsan Paik Hospital(Korea))
Keywords:
EEG,PPG,machine-learning,classification,computer-aided diagnosis (CAD)
The Maumgyeol platform is an AI-based mental health assessment system that utilizes objective biosignals such as electroencephalography (EEG) and photoplethysmography (PPG). It enables effective and location-independent monitoring of users' mental states. This study aimed to develop a machine learning-based diagnostic system for psychiatric disorders using wearable physiological signals collected through the Maumgyeol platform. EEG and PPG data were collected from 176 participants equally divided into four groups: major depressive disorder (MDD), anxiety disorder (ANX), schizophrenia (SCZ), and healthy controls (HC). Signals were acquired using the Maumgyeol Basic device, which records two-channel EEG and PPG-derived HRV data. After preprocessing, 22 features reflecting neural and autonomic activity were extracted and normalized. A RandomForest (RF) classifier was applied to classify the four groups. Model performance was evaluated using 10-fold cross-validation repeated 30 times. In each fold, features were ranked based on importance from the training data, and classification was performed using the top 1 to 22 features.The model achieved its highest average classification accuracy of 65.2% when using 20 features. Accuracy generally improved with an increasing number of features. Sensitivity and specificity for each class were as follows: MDD (65.7%, 92.9%), ANX (62.7%, 91.1%), SCZ (73.6%, 92.5%), and HC (58.9%, 93.0%). SCZ showed the highest sensitivity, while HC showed the lowest sensitivity, indicating that clinical groups were more easily distinguished than healthy controls. Some HC cases were misclassified as patient groups, reflecting the model's limited specificity for detecting non-clinical individuals.These findings suggest that EEG and PPG signals obtained from wearable devices can support multi-class classification of psychiatric conditions using machine learning. However, class-wise performance imbalance remains a challenge. Future research will focus on generating integrated EEG–PPG features and incorporating deep learning techniques to improve model performance and enhance clinical applicability.