Symptom Identification for Interpretable Detection of Multiple Mental Disorders on Social Media

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9970 - 9985 December 7-11, 2022 ©2022 Association for Computational Linguistics Symptom Identification for Interpretable Detection of Multiple Mental Disorders on Social Media Zhiling Zhang∗and Siyuan Chen∗ and Mengyue Wu† and Kenny Q. Zhu† Shanghai Jiao Tong University Shanghai, China {blmoistawinde, chensiyuan925, mengyuewu}@sjtu.edu.cn, kzhu@cs.sjtu.edu.cn Abstract Mental disease detection (MDD) from social media has suffered from poor generalizability and interpretability, due to lack of symptom modeling. This paper introduces PsySym, the first annotated symptom identification corpus of multiple psychiatric disorders, to facilitate further research progress. PsySym is anno- tated according to a knowledge graph of the 38 symptom classes related to 7 mental diseases complied from established clinical manuals and scales, and a novel annotation framework for diversity and quality. Experiments show that symptom-assisted MDD enabled by PsySym can outperform strong pure-text baselines. We also exhibit the convincing MDD explanations provided by symptom predictions with case studies, and point to their further potential ap- plications. 1 1 Introduction Mental health has been a significant challenge in global healthcare. Nearly 1 in 5 U.S. adults live with a mental illness or condition (NIMH, 2022), and there are about 1 billion people suffering from mental disorders worldwide (UN, 2020). Due to the stigma of mental disorders and lack of professional mental health services, many people cannot receive proper diagnose or treatment for their conditions. Social Media can be a promising source for studies on this problem, as we may detect hidden traces of mental disorders from the symptoms that users may reveal in their free sharing, and provide perti- nent help for those in need. Consequently, Mental Disease Detection (MDD) from social media has received increasing attention (Coppersmith et al., 2015; Cohan et al., 2018). However, most automatic MDD methods still struggle in this task, especially for their unsatis- ∗Equal Contribution †Corresponding Authors. 1Code at https://github.com/blmoistawinde/ EMNLP22-PsySym. Dataset can be provided upon request. problem with hoarding …. Disease Detection User Posts No Disease Diagnosis problem with hoarding …. Disease Detection OCD Symptom Identification Diagnostic Criteria for Obsessive-Compulsive Disorder (OCD): ✔ Obsession • Hoarding • contamination obsessions • …… ✔ Compulsions • …… ✔ Anxious Mood • …… Figure 1: Comparison between text-based and the proposed symptom-assisted mental disease detection method, which leverages psychiatric knowledge for symptom identification to improve the effectiveness and interpretability of MDD. fying generalizability and explainability. Firstly, these models may learn dataset-specific spurious correlations between certain words and the labels (usually diseases), and thus fail to generalize (Har- rigian et al., 2020). Moreover, most deep learning based methods work as black boxes, and cannot provide explanations for their prediction, which differs from the clinical practice which leverages symptom-based diagnostic criterions from author- itative manuals like DSM-5 (APA et al., 2013). Therefore, it may be hard for current MDD meth- ods to gain trust from their users. To tackle these issues, there has been a rising interest in utilizing symptoms for MDD, as they are the bases that …

Content truncated to 500 words