Diagnosis and Causal Analysis of Depression Based on EEG Features and Machine Learning

Depression is a serious mental illness that imposes a substantial burden on both individuals and society. There is a growing interest in employing learning-based methods utilizing bio-signals for the autonomic diagnosis of depression. However, the extent of bio-signal features' contributions in these models and the causal relationships among them remain underexplored. This study developed machine learning models for depression diagnosis utilizing electroencephalogram (EEG) data, with a focus on estimating the contributions of EEG features, including alpha interhemispheric asymmetry, activity, and mobility, within the learning-based frameworks. Additionally, the causal relationships among the most representative features were discussed. The experimental results showed that XGBoost had the best performance with the highest prediction accuracy among three learning-based models. Experimental findings also revealed that the interhemispheric asymmetry features provided the larger contributions in diagnosis of depression, than the features of activity and mobility. With three different causality analyses methods, the study identifies stable and tight causal relationships between the prefrontal and parietal asymmetry, and the depression. These findings provide a new perspective and basis for the efficient and reliable diagnosis of depression, which is helpful to improve the accuracy and efficiency.