Automated diagnosis and stage prediction of major depressive disorder using multimodal behavioral signals and correlation-integrated graph convolution network

One common and widespread mental health issue is major depressive disorder (MDD) that can have fatal consequences due to its effects on the body and mind. There is increasing interest in automating depression diagnosis and stage prediction using behavioral signals because of the dearth of diagnostic tools and subjectivity associated with diagnosing depression. The task is made more difficult by the vast range of conceivable behavior changes and the lack of labeled behavioral datasets for such problems. By choosing the most important characteristics from each modality to consider while making decisions, the Correlation-Integrated Graph Convolution Network (CIGCN) combines multi-level attention and strengthens overall learning. Empirical evidence shows that incorporating better correlation information is crucial for gaining greater accuracy in detecting depression in videos and text. The results of experiments conducted with the DAIC-WOZ dataset indicate that the combination of text-based video classification integrates two different modes of data. This integration is achieved through a model-level fusion approach, where the final representation of the GCN model is applied to a fully connected layer. This process allows for the estimation of both the PHQ score and binary level. By employing a multimodal strategy, we were able to attain an RMSE score of 3.26 on the test set and a binary classification F1 Score of 90.10%. This is a significant enhancement of 0.18 points compared to the baseline performance in the challenge.