Abstract
Deep learning techniques have demonstrated significant promise for detecting Major Depressive Disorder (MDD) from textual data but they still face limitations in real-world scenarios. Specifically, given the limited data availability, some efforts have resorted to aggregating data from different domains to expand the data volume. However, these approaches face critical challenges, including data privacy, domain gaps, class imbalance, and uncertainty arising from both the data and the model. To overcome these challenges, we propose an Uncertainty-Aware Domain Incremental Learning framework for Cross-Domain Depression Detection (UDIL-DD), integrating Uncertainty-guided Adaptive Class Threshold Learning (UACTL) and Data-Free Domain Alignment (DFDA). Specifically, our UACTL module measures the discrepancy between predictions across sequential domains and learns adaptive thresholds tailored to each class, incorporating predictive uncertainty to enhance robustness. Subsequently, the DFDA module leverages domain-similar samples identified by UACTL to approximate historical feature distributions without accessing previous domain data, effectively addressing catastrophic forgetting. To validate the effectiveness of the proposed method, we conduct extensive experiments on four benchmark MDD datasets-CMDC, DIAC-WoZ, MODMA and EATD confirming the effectiveness of our method's potential for reliable depression detection in real-world clinical scenarios.