Comparative Analysis of Feature Selection Methods for ERP based Depression Detection

Depression is silent yet serious illness leading to functional disabilities of individuals. Early diagnosis for depression is necessary to reduce socio-economic burden. Recently, electroencephalogram (EEG) based depression detection using machine learning classifiers has shown promising results. The classifier's performance depends upon many parameters and drastically changes with number and type of features fed to it. Study focuses on comparing performance analysis of feature selection methods for event-related potential (ERP) based depression detection. The publicly available dataset is used for the work. Major depressive disorder (MDD) patients and healthy controls (HC) are discriminated based on quantitative features extracted from ERP data. The diagnostic performance of ten feature selection methods is evaluated by using logistic regression (LR) classifier with 10-fold cross-validation (CV). The overall accuracy improvement with feature selection methods is observed as up to 6.63%. L1 (LASSO) regularization resulted as the best amongst all feature selection methods with 90.91% classification accuracy.