Elamur Research⌕
☰
home / papers / 2026 / Resource-Efficient LLMs for Depression Symptoms Screening: Performance and Limitations in Zero Shot Setting

Resource-Efficient LLMs for Depression Symptoms Screening: Performance and Limitations in Zero Shot Setting

Paper · 2026 · Muhammad Rizwan and Jure Demšar

Summary

The paper examines whether small and mid-sized open LLMs can recognise individual DSM-5 depression symptoms in short texts without training or examples. The authors used the ReDSM5 dataset of Reddit sentences, labelled by a licensed clinical psychologist for nine DSM-5 symptoms. They removed multi-label sentences and kept symptom-free sentences as a control class. They then prompted eight instruction-tuned models from the Llama 3.2, Mistral and Qwen 2.5 families to act as a clinician and output a single category. Decoding was deterministic. Qwen 2.5-7B performed best. Its 14B sibling did not improve on it, so larger size alone did not yield better symptom recognition. Error analysis showed that the model handles explicit signals such as suicidal ideation well. It often misreads depressed mood, missing irritability and social isolation, or mistaking ordinary sadness for a symptom. The work suggests modest open models could assist initial screening at low compute cost. However, without fine-tuning they risk missing clinically important symptoms and cannot replace professional assessment.

Key points

  • Data: ReDSM5 (Bao et al., 2025) was filtered to 1,610 single-label sentences in 10 classes: nine symptoms plus "None". Dropping multi-label items removed 129 sentences. The largest classes were None (374) and depressed mood (326), and the smallest was psychomotor (32).
  • Models and setup:
    • Models tested: Qwen 2.5 (1.5B, 3B, 7B, 14B), Llama 3.2 (1B, 3B) and Mistral (7B, 12B).
    • Inference: zero-shot prompting, temperature 0 and max_tokens 15.
    • Non-conforming outputs were discarded as invalid.
  • Best result: Qwen 2.5-7B reached Cohen's κ = 0.603 and Macro F1 = 0.648. Llama-3.2-1B-Instruct was close to chance (κ = 0.060, Macro F1 = 0.089).
  • Typical errors:
    • Depressed mood and None were frequently confused in both directions.
    • The model leaned on words like "sad" or "cry".
    • Neutral-sounding mentions of sleep hours were labelled as no symptom.
  • Reproducibility: The code is published on GitHub as rizwan2phd/zeroshot-depression-symptoms-screening-llms.

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

🎓University of LjubljanaLjubljana · Slovenia13 works · 38,693 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksPsychology · 1,649 worksSocial Psychology · 330 worksHealth Professions · 75 worksHealth Information Management · 56 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,894 works

Cite

Cite as Muhammad Rizwan and Jure Demšar. 2026. Resource-Efficient LLMs for Depression Symptoms Screening: Performance and Limitations in Zero Shot Setting. Proceedings of the Sixth Resources and ProcessIng of linguistic, para-linguistic and extra-linguistic Data from people with various forms of cognitive/psychiatric/developmental impairments in cooperation with the MENTAL.ai consortium.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/5fkbms3rp4bo

@inproceedings{rizwan2026resource,
  title = {{Resource-Efficient LLMs for Depression Symptoms Screening: Performance and Limitations in Zero Shot Setting}},
  author = {Muhammad Rizwan and Jure Demšar},
  booktitle = {Proceedings of the Sixth Resources and ProcessIng of linguistic, para-linguistic and extra-linguistic Data from people with various forms of cognitive/psychiatric/developmental impairments in cooperation with the MENTAL.ai consortium},
  year = {2026},
  doi = {10.63317/5fkbms3rp4bo},
  url = {https://doi.org/10.63317/5fkbms3rp4bo}
}

References

5 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Information Retrieval Lab, University of A Coruña. 2025. ReDSM5: A Reddit Dataset for DSM-5 Depression Detection. Information Retrieval Lab (IRLab), Universidade da Coruña. Hugging Face Datasets, Mental Health and Clinical NLP Resources, 1.0. PID https://huggingface.co/datasets/irlab-udc/redsm5. Red… ↗
  2. Gwenolé Loas, Guillaume Lefebvre, Marianne Rot-saert, and Yvon Englert. 2018. Relationships between anhedonia, suicidal ideation and suicide attempts in a large sample of physicians. PloS one, 13(3):e0193619. find
  3. Lydia Manikonda and Munmun De Choudhury. 2017. Modeling and understanding visual attributes of mental health disclosures in social media. In Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems, pages 170–181. find
  4. Mary L McHugh. 2012. Interrater reliability: the kappa statistic. Biochemia medica, 22(3):276– 282. find
  5. Mahmud Omar, Shelly Soffer, Alexander W Charney, Isotta Landi, Girish N Nadkarni, and Eyal find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/rizwan-2026-resource-efficient-llms-depression-symptoms/