Automatic Depression Detection Using Attention-Based Deep Multiple Instance Learning

Depression is a serious mental illness and one of the leading causes of suicide worldwide. However, the social prejudice and the lack of psychiatrists for depression lead to a significant number of depressed patients without accurate diagnosis and subsequent serious consequences. With the rise of social media, previous studies have found that the information of depressed patients on social media can be analyzed to automatically detect depression for auxiliary diagnosis. In the context of weakly supervised learning framework, a multiple instance learning (MIL) method is proposed to identify depression from social media with visual and vocal information. By leveraging the state-of-the-art attention-based deep LSTM (AD-LSTM), the proposed MIL method can handle the problem with sparse labels (i.e., one label for a long-term sequence of visual information). More specifically, the AD-LSTM module is used to process a fixed-length visual and vocal segments to extract temporal representations of instances, and the AD-MIL module is used to aggregate the obtained temporal representations for individual subject predictions. Compared with current benchmarks, our experiments demonstrate that our proposed MIL method can achieve the best weighted average precision, recall and F1 score with the corresponding values as 66.56%, 66.98% and 66.55%, respectively. The numerical results illustrate that the potential and effectiveness of our proposed MIL method in the field of depression detection.