0643 Machine Learning Model to Predict Depressive Mood Using Graph Theory Analysis of Polysomnography

Abstract

Abstract Introduction The authors conducted this preliminary study to discover the brain signal features for subjective depressive mood by the method that combines a machine learning approach and a graph theory analysis. Methods The authors retrospectively investigated the records of 13 subjects who had undertaken polysomnography (PSG) and patient health questionnaire-9 (PHQ-9). We divided the data to a training set (10 subjects) and a test set (2 subjects). The predictive modeling was conducted by 3 steps: 1) estimation of the small world coefficient (sigma, omega) of each frequency bands (delta: 0.5-4.0 Hz, theta: 4.0-8.0 Hz, alpha: 8.0-12.0 Hz, sigma: 12.0-16.0 Hz, beta: 16.0-30.0 Hz, gamma: 30.0-100.0 Hz) of electroencephalography (EEG) channel for N1 and N2 sleep, 2) training predictive models, and 3) evaluation of final performance of model by using test set. The authors used mean absolute error (MAE) and mean squared error (MSE) as evaluation metrics. The training was conducted using Python 3.8, and its libraries (MNE-Python, MNE-connectivity, NetworkX, xgboost scikit-learn). Results Median value of PHQ-9 score was 6.0 (interquartile range: 5, 8). Lasso regression model showed good performance (MAE 0.6, MSE 0.4). XGBoost model also showed good performance (MAE 0.8, MSE 1.3). Conclusion Lasso regression model using small world coefficient of EEG in overnight PSG showed good predictive value for PHQ-9. The further research using a sufficient sample is needed. Support (if any)