Comparing linear, nonlinear and time frequency-based features in the EEMD domain for depression detection application using EEG signals

Abstract

Depression is a disease that, if left untreated, can lead to risks such as self-harm and suicide. It can be treated with exercise and antidepressant drugs. Timely and accurate diagnosis of depressed subjects can help prevent such dangerous behaviors. An electroencephalogram (EEG) is an available tool in clinics and hospitals for measuring brain activity. In this article, the ability of
three kinds of parameters, including linear, nonlinear, and time-frequency-based features, was evaluated for the accurate detection of normal and depressed EEG signals. The EEG signals of 22 normal and 22 depressed subjects were used to evaluate the proposed framework. In the first step, the input EEG signals were decomposed into their intrinsic mode functions (IMFs) using the ensemble empirical mode decomposition (EEMD) method, and the best mode with the highest effect was selected by mutual information (MI). After that, these three kinds of features were extracted from the best IMF and fed to a k-nearest neighbors (KNN) classifier. We found that time-frequency features are better than nonlinear and linear features in detecting depressed EEG signals. Additionally, EEG signals from the right hemisphere of the brain were better than those from the left side for depression detection. In the final step of the proposed method, significant features were selected by the particle swarm optimization (PSO) algorithm and fed to the KNN
classifier, resulting in average classification accuracies (ACC) of 91% and 92% for the left and right hemispheres, respectively. The proposed method can be used in clinics and hospitals for accurate, fast, and accessible detection of depressed subjects.