Exploring Multi-Task Learning Paradigms for Detecting Mental Disorders Using Video Modality

Multi-Task learning (MTL) has been reported to enhance the performance of the task under consideration by exploiting knowledge from related tasks. A weighting method has been utilized to aggregate the MTL losses. The selection of a related task and a weighting method is vital for the performance of the task under consideration. These aspects have yet to be investigated in the literature for mental disorders detection tasks. In this study, we explore different related tasks and weighting methods for the detection of major depressive disorder (MDD) and post-traumatic stress disorder (PTSD) for video modality. The experimental results show that the MTL has outperformed the corresponding single-task learning (STL) for several combinations of related tasks and weighting methods. The best detection performance for MDD and PTSD is found to be 0.676 and 0.706, measured in the F1-score, and the relative increment compared to STL is 6.96% and 12.78%, respectively.