A text-based emotional pattern discrepancy aware model for enhanced generalization in depression detection

Text-based automated depression detection is one of the current hot topics. However, current research lacks the exploration of key verbal behaviors in depression detection scenarios, resulting in insufficient generalization performance of the models. To address this issue, we propose a depression detection method based on emotional pattern discrepancies, as the discrepancies are one of the fundamental features of depression as an affective disorder. Specifically, we propose an Emotional Pattern Discrepancy Aware Depression Detection Model (EPDAD). The EPDAD employs specially designed modules and loss functions to train the model. This approach enables the model to dynamically and comprehensively perceive the different emotional patterns reflected by depressed and healthy individuals in response to various emotional stimuli. As a result, it enhances the model’s ability to learn the essential features of depression. We evaluate the generalization performance of our model from a cross-dataset and cross-topic perspective using MODMA (52 samples) and MIDD (520 samples) datasets. In cross-topic generalization experiments, our method improves F1 score by 10.39% and 1.77% on MODMA and MIDD, respectively, in comparison to the state-of-the-art method. In cross-dataset generalization experiments, our method improves the F1 score by a maximum of 6.37%. We also compare our model with large language models, and the results indicate it is more effective for depression detection tasks. Our research contributes to the practical application of depression detection models. Our code is available at: https://github.com/hbZhzzz/EPDAD.