The research, published in BMC Psychiatry, enrolled 108 adolescents and young patients diagnosed with major depressive disorder according to DSM-5-TR criteria, all receiving outpatient care. The participants had a mean age of 15.8 years, and 58 of them, or 53.7 percent, were female. The team, led by researchers at Wonkwang University and Wonkwang University Hospital with collaborators at Seoul National University Hospital and NAVER, set out to answer a deceptively simple question: can the everyday narratives patients generate about their own lives add meaningful signal to the standardized clinical measures psychiatrists already collect?

The study’s architecture reflected that dual-source philosophy. Structured instruments assessed suicidal ideation, depressive symptoms, anxiety, and perceived stress, including the Scale for Suicide Ideation and child and adolescent depression scales. Alongside these, participants kept diary entries, producing 709 valid entries across the cohort that were used for the sensitive-topic evaluation. The self-report outcome the researchers targeted was positivity on item 9 of the Patient Health Questionnaire, which asks about thoughts of being better off dead or of hurting oneself; a score of 1 to 3 counted as positive, while a 0 counted as negative.

In patient-level logistic regression, two structured measures stood out. Each point increase on the Scale for Suicide Ideation raised the odds of PHQ-9 item 9 positivity by 15 percent, with an odds ratio of 1.15 and a 95 percent confidence interval of 1.02 to 1.28. Depressive symptoms also carried signal, with an odds ratio of 1.09 per point and a confidence interval of 1.00 to 1.19. These findings confirmed that conventional clinical indicators do track self-reported suicidal thinking, but the researchers suspected they were incomplete on their own.

The decisive test came when diary narratives were added to the classification pipeline. Two large language models, GPT-4o and DeepSeek-V3, were benchmarked against blinded mental health professionals on cross-sectional classification tasks. For GPT-4o, adding diary narratives to depressive symptom scores improved the F1-score, a harmonic mean of precision and recall, from 0.79 to 0.81. DeepSeek-V3 showed a larger gain, rising from 0.74 to 0.78. In the primary aggregate longitudinal analysis, classification performance increased as the number of diary entries per patient grew, and a sensitivity analysis restricted to the same 16 participants across diary counts showed positive associations for both models and for the human professionals alike.

Perhaps the most technically demanding component was sensitive-topic detection within the journals themselves. Clinicians annotated entries that contained concerning content, and the language models were evaluated against those human labels. The LLM detection achieved an F1-score of 0.85 on clinician-annotated in-journaling sensitive topics, a level of agreement that surprised even the study team given how subtle, metaphorical, and context-dependent adolescent expressions of distress can be. The models were not simply keyword matching; they were integrating narrative context that structured questionnaires cannot capture.

These components were then woven into an integrated workflow called MindfulDiary, which combined three layers of protection: pre-journaling screening, in-journaling LLM detection, and clinician alerting. On the composite outcome, the workflow captured 193 of 197 composite-positive entries, a recall of 98.0 percent, with an overall F1-score of 0.96. Critically, the system was designed as a human-in-the-loop pipeline. No alert ever reached a clinical decision without a clinician reviewing it, and the authors are explicit that the tool is not an autonomous suicide-risk classifier.

That caution matters. Suicide risk assessment in minors is among the highest-stakes applications in clinical artificial intelligence, and false positives can burden families and clinicians while false negatives can be catastrophic. The study’s observational, cross-sectional design, with longitudinal analyses that remain exploratory, means the results demonstrate feasibility and validity of the signal rather than proven clinical benefit. The authors note that evaluation across multiple sites and languages, prospective studies, and patient-centered evaluation are all needed before any routine adoption. The study was also conducted at a single healthcare system in Korea, and diary writing behavior may differ across cultures and age groups.

The ethical scaffolding of the research was correspondingly rigorous. The study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Review Board of Wonkwang University Hospital, and informed consent was secured from at least one legal guardian for every participant. The design of the MindfulDiary workflow drew on prior field experience with psychiatric patients and clinicians, although the authors acknowledge that no patient or public representatives were involved in manuscript preparation or dissemination planning. The work was funded by Korean government agencies including the Korea Health Industry Development Institute, the National Research Foundation of Korea, and the Ministry of Education.

What makes the study resonate beyond psychiatry is its reframing of what patient-generated data can mean. For decades, the quantitative ideal in medicine has been the standardized questionnaire, and narratives were treated as anecdote. This research shows that free-text narratives and structured indicators are complementary rather than competing: the scales anchor the measurement in validated psychometrics, while the diaries reveal the fluctuating daily emotional context that brief outpatient encounters systematically miss. As large language models mature, the marginal value of listening to what patients voluntarily write, with appropriate safeguards, may prove to be one of the most clinically consequential applications of the technology. For now, the message is measured: AI can help clinicians keep watch between visits, but the human in the loop remains indispensable.

Subject of Research: LLM-assisted journaling for monitoring suicidal ideation in adolescents with major depressive disorder

Article Title: Integrating patient narratives and clinical indicators for monitoring suicidal ideation in adolescents with major depressive disorder: a human-in-the-loop validation study of LLM-assisted journaling

Article References: Kim, J.-W., Yoon, H., Oh, W., Jung, D., Jo, Y., Lee, S.-W., Jeong, C.-K., Yoon, S.-H., Kim, D.-J., Jung, S.-I., Lee, S.-Y., Kim, B.-N., Yoo, H., Kim, Y.-H., & Yang, C.-M. (2026). Integrating patient narratives and clinical indicators for monitoring suicidal ideation in adolescents with major depressive disorder: a human-in-the-loop validation study of LLM-assisted journaling. BMC Psychiatry. https://doi.org/10.1186/s12888-026-08700-y

Image Credits: AI Generated

DOI: 10.1186/s12888-026-08700-y

Keywords: adolescent mental health, major depressive disorder, suicidal ideation, large language models, digital mental health, journaling, human-in-the-loop, clinical decision support, patient-generated narratives, psychiatry, PHQ-9, MindfulDiary