AMG: A Depression Detection Model with Autoencoder and Multi-Head Graph Convolutional Network

Depression is a common chronic mental disorder characterized by high prevalence, recurrence, suicide rate, disability rate and heavy disease burden. Depression assessment based on questionnaires is subjective, therefore, a novel framework to is hereby designed to detect depression automatically and obj ectively. The framework first utilizes a two-way autoencoder to extract the hidden features from the original differential entropy features. The new features are used to generate the channel connectivity graph, using the multi-head attention mechanism. Next, the new features and graphs are input into the graph convolutional network (GCN) to aggregate spatial information. Finally, the output from GCN is passed to a fully connected layer and a soft layer to obtain the predicted label. Experiments are performed on Multimodal Open Dataset for Mental-disorder Analysis dataset. A final accuracy of 88.68% and a cross entropy loss of 2.4532 are obtained, when using ten-fold cross validation, outperforming other traditional machine learning methods including SVM, KNN, decision tree.