Elamur Research⌕
☰
home / papers / 2026 / Rethinking Depression Prediction from a Fine-Grained Subscore Modeling Perspective via Multi-Task Learning

Rethinking Depression Prediction from a Fine-Grained Subscore Modeling Perspective via Multi-Task Learning

Paper · 2026 · Zhenguang Wang, Bo Li, Wenhui Tan et al.

Summary

Automated depression assessment usually reduces a clinical interview to a single number, such as a total scale score or a severity class. Yet people with equal totals can have very different symptom profiles. The authors show on clinical data that this variation is systematic and strongest for totals between 12 and 20. They propose MTSP, which treats each item of a rating scale as its own regression task inside one multi-task model. Transcripts are split into utterances and encoded with MentalBERT (English) or MedBERT-Chinese. The encodings then feed a shared task encoder and item-specific heads. A clinically informed task correlation graph penalizes divergent predictions for items in the same symptom cluster. Task-level self-paced learning reweights item losses by how hard each item is. Experiments cover PHQ-8 on E-DAIC and HAMD-13 on the Chinese PDCH and private CIDH sets. MTSP outperforms total-score regression and zero-shot Qwen3-14B scoring, with the best reported MAE on E-DAIC. Its representations group patients by symptom pattern rather than by total alone, which supports more interpretable clinical decision support.

Key points

  • On E-DAIC (PHQ-8, 8 items scored 0–3), MTSP reaches a state-of-the-art MAE of 3.48. Its RMSE of 4.57 is on par with the strongest earlier system.
  • HAMD-13 uses only the 13 text-assessable items of HAMD-17. It leaves out Depressed Mood, Retardation, Agitation and Depersonalization, which need behavioral observation.
  • The data consist of CIDH, a private clinical set of 1,689 interviews, PDCH, a small public set of 99 samples, and E-DAIC, which uses virtual-interviewer sessions.
  • Ablations indicate three things: multi-task learning is necessary, subscore supervision captures symptom-cluster structure, and the prior constraints plus T-SPL improve robustness to uneven item difficulty and label noise. Qwen3-14B zero-shot shows per-item MAE ranging from 0.57 to 1.05 on PHQ-8.
  • In t-SNE plots, subscore training places two CIDH patients who both total 15 in separate regions. It also gives higher pattern coherence than total-score training (0.780 vs 0.742). Code is released at github.com/wxcwzg/MTSP.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Renmin University of ChinaBeijing · China114 works · 16,118 citations🎓University of AlbertaEdmonton · Canada44 works · 3,236 citations🎓Northeastern UniversityShenyang · China42 works · 1,940 citations

Also on the byline: Nanjing Medical University · Nanjing Brain Hospital.

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksPsychology · 1,552 worksExperimental and Cognitive Psychology · 870 worksSocial Psychology · 306 works

Cite

Cite as Zhenguang Wang, Bo Li, Wenhui Tan, Peng Cao, Yang Wang, Jia Duan, Fei Wang and Osmar Zaiane. 2026. Rethinking Depression Prediction from a Fine-Grained Subscore Modeling Perspective via Multi-Task Learning. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.1841

@inproceedings{wang2026rethinking,
  title = {{Rethinking Depression Prediction from a Fine-Grained Subscore Modeling Perspective via Multi-Task Learning}},
  author = {Zhenguang Wang, Bo Li, Wenhui Tan, Peng Cao, Yang Wang, Jia Duan, Fei Wang and Osmar Zaiane},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.1841},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1841}
}

References

33 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Navneet Agarwal, Kirill Milintsevich, Lucie Metivier, Maud Rotharmel, Gaël Dias, and Sonia Dollfus. 2024. Analyzing symptom-based depression level estimation through the prism of psychiatric expertise. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language R… find
  2. Robert Michael Bagby, Andrew G. Ryder, Deborah R. Schuller, and Margarita B. Marshall. 2004. The hamilton depression rating scale: has the gold standard become a lead weight? The American journal of psychiatry, 161 12:2163–77. find
  3. Guanqun Bi, Zhuang Chen, Zhou Liu, Hongkai Wang, Xiyao Xiao, Yuqiang Xie, Wen Zhang, Yongkang Huang, Yuxuan Chen, Libiao Peng, Yi Feng, and Minlie Huang. 2025. Magi: Multi-agent guided interview for psychiatric assessment. In Annual Meeting of the Association for Computational Linguistics. find
  4. Amanda M. Buch and Conor Liston. 2020. Dissecting diagnostic heterogeneity in depression by integrating neuroimaging and genetics. Neuropsychopharmacology, 46:156–175. find
  5. Isobel M. Cameron, John R. Crawford, Kenneth Lawton, and Ian Cameron Reid. 2008. Psychometric comparison of phq-9 and hads for measuring depression severity in primary care. The British journal of general practice : the journal of the Royal College of General Practitioners, 58 546:32–6. find
  6. Jianning Li, Amin Dada, Behrus Puladi, Jens Kleesiek, and Jan Egger. 2023. Chatgpt in healthcare: A taxonomy and systematic review. Computer methods and programs in biomedicine, 245:108013. find
  7. Pengfei Cao, Yuanzhe Zhang, Chenxiang Zhang, Wei Chen, Yan Liu, Shuang Xu, Miao Xu, Wenqing Jin, Jinjie Xu, Dan Wang, Wei Wang, Xue Wang, Wen Wang, Yan ping Ren, Jun Zhao, Rena Li, and Kang Liu. 2025. A multimodal depression consultation dataset of speech and text with hamd-17 assessments. Scientifi… find
  8. Aishik Mandal, Dana Atzil-Slonim, Thamar Solorio, and Iryna Gurevych. 2025. Enhancing depression detection via question-wise modality fusion. ArXiv, abs/2503.20496. find
  9. Munmun De Choudhury, Michael Gamon, Scott Counts, and Eric Horvitz. 2013. Predicting depression via social media. Proceedings of the International AAAI Conference on Web and Social Media. find
  10. Glen A. Coppersmith, Mark Dredze, and Craig Harman. 2014. Quantifying mental health signals in twitter. In CLPsych@ACL.
  11. Chengfeng Dou, Chong Liu, Fan Yang, Fei Li, Jiyuan Jia, Mingyang Chen, Qiang Ju, Shuai Wang, Shunya Dang, Tianpeng Li, Xiangrong Zeng, Yijie Zhou, Chenzheng Zhu, Dawei Pan, Fei Deng, Guangwei Ai, Guosheng Dong, Hongda Zhang, Jinyang Tai, and 14 others. 2025. Baichuan-m2: Scaling medical capability w…
  12. Robert D. Gibbons, David C. Clark, and David J. Kupfer. 1993. Exactly what does the hamilton depression rating scale measure? Journal of psychiatric research, 27 3:259–73. find
  13. George Gkotsis, Anika Oellrich, Sumithra Velupillai, Maria Liakata, Tim J. P. Hubbard, Richard J. B. Dobson, and Rina Dutta. 2017. Characterisation of mental health conditions in social media using informed deep learning. Scientific Reports, 7. find
  14. Max Hamilton. 1960. A rating scale for depression. Journal of Neurology, Neurosurgery & Psychiatry, 23:56 – 62. find
  15. Shaoxiong Ji, Tianlin Zhang, Luna Ansari, Jie Fu, Prayag Tiwari, and E. Cambria. 2021. Mentalbert: Publicly available pretrained language models for mental healthcare. ArXiv, abs/2110.15621. find
  16. Kurt Kroenke, Robert L. Spitzer, and Janet B W Williams. 2001. The phq-9: validity of a brief depression severity measure. Journal of general internal medicine, 16 9:606–13. find
  17. Kurt Kroenke, Tara W. Strine, Robert L. Spitzer, Janet B W Williams, Joyce T. Berry, and Ali H. Mokdad. 2009. The phq-8 as a measure of current depression in the general population. Journal of affective disorders, 114 1-3:163–73. find
  18. M. Pawan Kumar, Ben Packer, and Daphne Koller. 2010. Self-paced learning for latent variable models. In Neural Information Processing Systems. find
  19. Kirill Milintsevich, Kairit Sirts, and Gael Dias. 2023. Towards automatic text-based estimation of depression through symptom prediction. Brain Informatics, 10. find
  20. Syed Arbaaz Qureshi, G. Dias, Mohammed Hasanuzzaman, and Sriparna Saha. 2020. Improving depression level estimation by concurrently learning emotion intensity. IEEE Computational Intelligence Magazine, 15:47–59. find
  21. Anupama Ray, Siddharth Krishna Kumar, R. Venkata Siva Reddy, Prerana Mukherjee, and Ritu Garg. 2019. Multi-level attention network using text, audio and video for depression prediction. Proceedings of the 9th International on Audio/Visual Emotion Challenge and Workshop. find
  22. Fabien Ringeval, Björn Schuller, Michel F. Valstar, Nicholas Cummins, Roddy Cowie, Leili Tavabi, Maximilian Schmitt, Sina Alisamir, Shahin Amiriparian, Eva-Maria Messner, Siyang Song, Shuo Liu, Ziping Zhao, Adria Mallol-Ragolta, Zhao Ren, M. Soleymani, and Maja Pantic. 2019. find
  23. Avec 2019 workshop and challenge: State-of-mind, detecting depression with ai, and cross-cultural affect recognition. Proceedings of the 9th International on Audio/Visual Emotion Challenge and Workshop. find
  24. Misha Sadeghi, Bernhard Egger, Reza Agahi, Robert Richer, Klara Capito, Lydia Helene Rupp, Lena Schindler-Gmelch, Matthias Berking, and Bjoern M. Eskofier. 2023. Exploring the capabilities of a language model-only approach for depression detection in text data. 2023 IEEE EMBS International Conferenc… find
  25. Misha Sadeghi, Robert Richer, Bernhard Egger, Lena Schindler-Gmelch, Lydia Helene Rupp, Farnaz Rahimi, Matthias Berking, and Bjoern M. Eskofier. 2024. Harnessing multimodal approaches for depression detection using large language models and facial expressions. NPJ Mental Health Research, 3. find
  26. Fabian Schmidt, Seyedehmoniba Ravan, and Vladimir Vlassov. 2025. Probabilistic textual time series depression detection. ArXiv, abs/2511.04476. — [arXiv:2511.04476](https://arxiv.org/abs/2511.04476) find
  27. Floris Van Steijn, Gizem Sogancioglu, and Heysem Kaya. 2022. Text-based interpretable depression severity modeling via symptom predictions. Proceedings of the 2022 International Conference on Multimodal Interaction. find
  28. Laurens van der Maaten and Geoffrey E. Hinton. 2008. Visualizing data using t-sne. Journal of Machine Learning Research, 9:2579–2605. find
  29. Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. 2017. Graph attention networks. ArXiv, abs/1710.10903.
  30. World Health Organization. 2025. Depressive disorder (depression). https://www.who.int/news-room/ fact-sheets/detail/depression. Accessed: 2025-12-15. ↗
  31. Danai Xezonaki, Georgios Paraskevopoulos, Alexandros Potamianos, and Shrikanth Narayanan. 2020. Affective Conditioning on Hierarchical Attention Networks Applied to Depression Detection from Transcribed Clinical Interviews. In Interspeech 2020, pages 4556–4560. find
  32. Feihong Yang, Xuwen Wang, and Jiao Li. 2021. Exploration and research on applying BERT to chinese clinical natural language processing. GitHub repository. Original title in Chinese: “BERT”. Accessed: 2026-01-04. find
  33. Linhai Zhang, Ziyang Gao, Deyu Zhou, and Yulan He. 2025. Explainable depression detection in clinical interviews with personalized retrieval-augmented generation. In Annual Meeting of the Association for Computational Linguistics. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wang-2026-rethinking-depression-prediction-fine-grained/