An Instant Depression Screening Method via Valence and Arousal Prediction from Electroencephalogram and Galvanic Skin Response Data via Unsupervised Representation Learning

Abstract

Globally, 280 million are suffering from depressive disorders, one of the most common mental disorders referring to long periods of depressed mood that affect life negatively, often leading to suicide. Adolescents are especially vulnerable, with 35% of youth aged 12 to 17 in the United States having major or severe depression. The rate of depression in teenagers has doubled over the past decade, along with the youth suicide rate. Currently, self-report questionnaires such as the Patient Health Questionnaire are utilized to diagnose depression. Self-report questionnaires may be accurate for other population groups that are fully aware of the state of their mental health, and are willing to seek support. However, adolescents do not have the same ability to recognize their symptoms nor answer honestly in school-wide screening tests. In order to solve the aforementioned problem, a method that captures the true emotion without self-reports based on electroencephalogram (EEG) and galvanic skin response (GSR) is proposed. The novel system is composed of representation and transfer learning steps for the extraction of emotion-related features from EEG signals, and a 1D convolutional neural network for extraction in GSR signals. The extracted features are merged and outputted as points on the arousal-valence graph, where emotions can be detected. Through extensive experiments, the proposed model demonstrated exceptional performance. The best MAE was 18.4 when without GSR, 18.8 when without representation learning, and 17.2 when without the proposed equation, but improved to 16.4 in the proposed model.