Detecting Depression, Anxiety and Mental Stress in One Sequential Model with Multi-task Learning

Shen Zhang, Mei Tu, Yueru Yan, Yimeng Zhuang, Likun Ge, Gaoxia Wei

Abstract: Recent studies employing mobile data to detect psychological disorders have made great progress. Mobile devices record user’s behavior data and physical signals continuously. Such data are able to detect symptoms of mental disorders like depression, anxiety and stress via machine learning models. However, existing researches mainly focus on independently detecting one of these disorders, ignoring the symptom relationship among them. Besides, current studies using statistical features accumulated in a period of time lose the time correlation within data, which makes it difficult to predict future symptoms. Instead, in this paper, we firstly propose a parameter sharing model with multi-task learning to transfer the common representation of three symptoms of mental disorders, and we use biRNN approach to predict these disorders by using mobile sequential features. We obtain 175 participants completing collection for at least 2 weeks with weekly questionnaire. In the experiments, our proposed sharing model achieves average overall accuracy of 0.78 and average AUC of 0.78, outperforming three single-task model and machine learning methods that use statistical features significantly. These results suggest that multi-task learning with the sequential feature enables detecting severity of depression, anxiety and stress symptoms.