EEG-Based Depression Detection with a Synthesis-Based Data Augmentation Strategy

Recently, Electroencephalography (EEG) is wildly used in depression detection. Researchers have successfully used machine learning methods to build depression detection models based on EEG signals. However, the scarcity of samples and individual differences in EEG signals limit the generalization performance of machine learning models. This study proposed a synthesis-based data augmentation strategy to improve the diversity of raw EEG signals and train more robust classifiers for depression detection. Firstly, we use the determinantal point processes (DPP) sampling method to investigate the individual differences of the raw EEG signals and generate a more diverse subset of subjects. Then we apply the empirical mode decomposition (EMD) method on the subset and mix the intrinsic mode functions (IMFs) to synthesize augmented EEG signals under the guidance of diversity of subjects. Experimental results show that compared with the traditional signal synthesis methods, the classification accuracy of our method can reach 75% which substantially improve the generalization performance of classifiers for depression detection. And DPP sampling yields relatively higher classification accuracy compared to prevailing approaches.