EMAA: An Ensemble Multimodal Approach based on Auto-decoder for Depression Detection

Depression has become a major disease in recent years, causing many people to commit suicide. The use of non-invasive electroencephalography (EEG) is crucial for the automated detection of depressive disorders. However, the current accuracies of EEG-based diagnostic methods need to be improved for real-world applications. To address this challenge, we used Common Spatial Patterns (CSP), a common technique in brain-computer interfaces (BCI), to extract features from event-related potential data and classify depression/health instances. We proposed the Ensemble Multimodal Approach based on Auto-decoder (EMAA) to promote depression recognition accuracy. According to the feature subsets generated by text matching, we then extract the feature vectors that exhibit outstanding performance in multimodal feature fusion from their corresponding subsets. Finally, the LightGBM classifier utilized feature vectors as input to classify the participants into their respective classes. The results showed that severely depressed and healthy participants were recognised with an accuracy of 91.34% through the fusion of happy and sad subsets using EMAA. In this work, we presented an automated processing method that could be a valuable reference for studies aimed at detecting EEG-based mild depression detection.