Social and Non-social Reward Learning Contexts for Detection of Major Depressive Disorder Using EEG: A Machine Learning Approach

Philopateer Ghattas, Mai Gamal, Seif Eldawlatly

Abstract: Major Depressive Disorder (MDD) is a leading cause of disability globally and a major cause of suicide deaths. Improving our understanding of MDD is expected to inspire better objective diagnostic and treatment tools which may decrease the burden of disease worldwide. MDD is associated with social impairments and reward processing aberrations. However, the interaction between social cognition and reward learning in MDD is not well understood. In this work, we aim to study the effect of integrating social information with reward learning in MDD using an EEG-based machine learning approach. We recorded EEG data from subjects during their participation in a reward learning experiment in social and non-social contexts. We then extracted linear and nonlinear features from the EEG data to detect MDD using Support Vector Machine (SVM) and K-Nearest Neighbors (KNN) classifiers. Our results show that the data collected during social contexts achieved broadly higher classification accuracies compared to non-social contexts reaching 80% using multi-channels. Moreover, single-channel data achieved comparable and even better accuracies than multi-channels with also superior performance in social contexts reaching 85.7%. In the sub-band classification analysis, beta and alpha bands were considerably better than theta and delta bands. Surprisingly, the non-social context had the highest accuracy in the beta band while the social context had the highest accuracy in the alpha band. The results were consistent when using single features compared to combining the features. These findings show the notable role of social information with reward processing in advancing our understanding of MDD and its subtypes which can guide the improvement of diagnostic and personalized treatment tools.