An Audio Correlation-Based Graph Neural Network for Depression Recognition

Depression is a prevalent mental health disorder. The diagnosis of depression hinges largely on the medical practitioner’s subjective assessment of the patient’s diagnostic process. The involvement of multiple subjective factors during this process can further complicate the diagnosis. In this paper, we propose a novel approach for depression recognition using the graph neural network that incorporates potential connections within and between audio signals. Specifically, we first extract time series information between frame-level audio signal features through GRU. We then construct two graph neural network modules to explore the potential connections of inter-audio and inter-audio. In the first graph module, we construct a graph using the frame-level features of each audio as nodes and embed the output graph into a feature vector representation. In the second graph module, we represent the graph embedding feature vector as a node and encode the potential relationships between audio signals through node neighbourhood information propagation. Additionally, we extract emotional features related to depression using a pre-trained emotion recognition network and enhance the connection between coded audio signals through a self-attention mechanism to further improve the model’s performance. We conducted extensive experiments on three depression datasets, and our proposed model outperformed all benchmark models, demonstrating its effectiveness.