A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies

Abstract Background AI technologies have considerable potential to support mental health practitioners in identifying depression symptoms. However, existing tools are often developed with limited practitioner involvement and evaluated primarily on technical performance, leaving questions of clinical relevance, ethical design, and practitioner trust insufficiently addressed. Objective This study aimed to examine how iterative co-design with mental health practitioners can inform the development of an AI-based model for identifying depression symptoms, focusing on how practitioner input shaped its clinical relevance and ethical foundations across 2 design cycles. Method We conducted a 2-stage qualitative co-design study. In stage 1, we developed DepressionSymp 1.0, a fine-tuned GPT-3.5 model trained on the CounselChat dataset, identifying 5 symptom categories based on the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition ( DSM-5 ) criteria and the Linguistic Inquiry and Word Count (LIWC-22) psycholinguistic framework, reviewed and discussed by 6 mental health practitioners during online co-design workshops. In stage 2, this feedback informed DepressionSymp 2.0, expanded to 10 symptom categories and reviewed by 8 further practitioners in a second round of workshops. Results Practitioners valued the structured symptom outputs for reducing cognitive workload and supporting intervention planning. In stage 1, indirect indicators such as social withdrawal, financial difficulties, and major life changes were often classified as lifestyle factors rather than symptoms; this informed the expansion to 10 categories in stage 2, described as clearer and more useful for preassessment. Risk categorization was identified as the most ethically sensitive aspect across both stages, alongside broader concerns including the lack of nonverbal information in text-based assessments, cultural and linguistic differences in symptom expression, and the potential for automation bias without sufficient human oversight. Conclusions Iterative co-design with mental health practitioners is valuable for developing AI-based tools that are clinically relevant and ethically grounded, demonstrating how qualitative practitioner feedback can inform iterative model refinement and ethical design. Findings inform design implications around large language model (LLM) features, practitioner-in-the-loop oversight, and ethical considerations covering clinical standards, informed consent, and data governance. Future work should prioritize larger-scale independent validation, formal safety evaluation, and culturally diverse dataset development.