Feature Selection and Imbalanced Data Handling for Depression Detection

Major Depressive Disorder (MDD) is the most common disorder worldwide. Accurate detection of depression is a challenging problem. Machine learning-based automated depression detection provides useful assistance to the clinicians for effective depression diagnosis. One of the most fundamental steps in any automated detection is feature selection and investigation of the most relevant features. Studies show that regional volumes of the brain are affected in response to depression. Regional volumes are considered as features. The gray matter volumes’ correlation with depression and the most effective gray volumes for depression detection is investigated in this study. Various feature selection techniques are studied, along with the investigation on the importance of resampling to handle imbalanced data, which is typically the case for depression detection, as the number of depressed instances is commonly a fraction of the entire data size. Experimental results using Random Forests (RF) and support vector machines (SVM) with a Gaussian kernel (RBF) as classifiers show that feature selection followed by data resampling gives superior performance measured by Area Under the ROC Curve (AUC) as well as prediction accuracy, and RF outperforms SVM for the depression detection task.