Interpretable depressive symptoms screening via statistical reasoning-augmented large language models using wearable...

Early assessment of depressive symptoms is essential for scalable and personalized mental health care. We developed a hybrid clinical decision support system (CDSS) that combines interpretable logistic regression with fine-tuned large language models (LLMs) to classify depressive symptom status using wearable-derived behavior, clinical features, and environmental exposures. This study aimed to develop a hybrid clinical decision support system by fine‑tuning a GPT‑based large language model with odds ratio-derived statistical reasoning to classify depressive symptom status from wearable‑derived behavioral metrics, clinical characteristics, and environmental exposures. We compared the diagnostic accuracy, interpretability, and operational efficiency of this OR‑based fine‑tuned model against both traditional nomogram‑based statistical models and label‑supervised fine‑tuned LLM approaches. We further evaluated the incremental predictive value of wearable‑derived behavioral features beyond readily available demographic and clinical variables using matched feature-set comparisons. We analyzed data from 2437 adults at Chonnam National University Hospital, South Korea (August 2021-January 2023) who wore wrist actigraphy devices continuously for 7 days. Depressive symptoms were assessed using PHQ-9, with a score ≥ 10 indicating higher symptom burden. We extracted temporal behavioral features including morning steps (06:00-12:00), non-daylight steps, and vigorous physical activity minutes, stratified by weekday/weekend. Meteorological data (temperature, precipitation) were integrated with behavioral metrics. Multivariable logistic regression identified significant predictors with odds ratios (ORs). Five LLMs (GPT-4o, o3-mini, Claude 3.5-sonnet, Gemini 2.0-flash, Llama 3.3) were evaluated under zero-shot conditions. The GPT-based model was selected for supervised fine-tuning using two strategies: label-based (binary outcome only) and OR-based (embedding statistical reasoning into prompts). The system was deployed as a chatbot interface. Among 2437 participants (mean age 60.2 years, 67.5% female), 317 (13.0%) were classified as having elevated depressive symptoms (PHQ-9 ≥ 10). Prespecified incremental-value analyses using matched feature sets (clinical-only, wearable-only, and combined) demonstrated that wearable-derived behavioral features provided additive predictive value beyond demographic and clinical variables. In the OR-based fine-tuned LLM framework, balanced accuracy increased from 0.635 (clinical-only) to 0.735 (wearable-only) to 0.787 (combined) on weekdays, and from 0.589 to 0.647 to 0.680 on weekends. In the nomogram framework, adding wearable features to the clinical model significantly improved risk reclassification (continuous NRI: 0.469, p < 0.001 on weekdays; 0.372, p = 0.002 on weekends) and integrated discrimination improvement (IDI: 0.037, p < 0.001 on weekdays; 0.018, p = 0.002 on weekends). The combined OR-based fine-tuned model achieved 90.3% accuracy and 94.5% specificity on weekdays and 88.8% accuracy and 96.4% specificity on weekends, outperforming both nomogram and label-based fine-tuned models. This hybrid framework demonstrates that odds ratio-anchored fine-tuning can improve classification performance while preserving predictor traceability for PHQ-9-defined elevated depressive symptoms. Wearable-derived behavioral features provided statistically significant incremental predictive value beyond demographic and clinical variables, with larger gains observed on weekdays than on weekends. Prospective external validation across diverse populations is required before clinical implementation.