Truace
Subscribe

Assisting Screening of Adolescent Depression With Large Language Models as Symptom Detectors: Algorithm Development and Validation

Background: Depression prevalence is rising in adolescents and young adults (ages 10-24 years). Traditional screening methods for depression, such as the Patient Health Questionnaire (PHQ-9), are particularly challenging for adolescents. AI has potential to help with this, but the scarcity of annotated datasets highlights the need for off-the-shelf approaches. In this study, we investigate the feasibility of using state-of-the-art large language models (LLMs) to detect depressive symptoms in adolescent settings…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-2207· cite & verify
Assisting Screening of Adolescent Depression With Large Language Models as Symptom Detectors: Algorithm Development and Validation
Indian health service [electronic resource] : a comprehensive health care program for American Indians and Alaska natives by United States. Indian Health Service. Public domain

In brief

Researchers tested three language models, Llama, FLAN-T5 and Phi, on free-text electronic health records from adolescents aged 15 to 17 with depression or related mood disorders at Cincinnati Children's Hospital Medical Center. The models were prompted without extra training to spot symptom categories drawn from the PHQ-9 and Beck Depression Inventory, and were compared with a simple word-search method.

The work suggests that existing models could support adolescent depression screening where annotated datasets are scarce, and that a small model like FLAN-T5 might suit resource-limited clinics. However, the manual annotation covered only 22 patients, the model comparisons were not standardized, and the authors say the findings need validation in broader populations and other mental health conditions.

Main points

  1. The study drew on records of 924 adolescent patients aged 15 to 17, but manual annotation covered only 22 patients across 16 symptom categories.
  2. FLAN-T5 reached average F1 of 0.65 and precision of 0.78, and scored 0.92 on sleep problems and 0.80 on self-loathing.
  3. Llama 3 had the highest recall at 0.90 but overgeneralized, which makes it less suitable for precise screening.
  4. Features from FLAN-T5 annotations fed a machine learning model that separated depression cases from controls with precision of 0.78.

The gain

In a pilot on adolescent health records, off-the-shelf language models found depressive symptoms in clinical notes far better than a word-search baseline, with FLAN-T5 giving the most precise results.

The rundown

The authors note that traditional screening tools such as the PHQ-9 are especially hard to use with adolescents, while depression prevalence is rising among people aged 10 to 24. They argue that scarce annotated datasets make off-the-shelf approaches attractive.

The three models were run in zero- and few-shot settings, so no task-specific training was needed. Phi gave a more balanced result, with precision of 0.44 and recall of 0.60, while FLAN-T5 offered the most favorable balance between precision and recall within the configurations evaluated.

The authors say computational efficiency of FLAN-T5 supports possible deployment in resource-limited clinical settings, and they call for standardized prompting strategies and task formulations to allow more definitive comparisons between model architectures.

Sources

  1. Peer-reviewedJMIR Formative Research2026-10-21

The debate