Healthcare AI & Clinical Decision Support

Early and accurate assessment of depressive symptoms is crucial for effective mental healthcare. Traditional methods often rely on retrospective self-reporting, which can miss dynamic fluctuations. This study addresses these limitations by developing a hybrid clinical decision support system (CDSS) that integrates diverse data sources and advanced AI techniques.

The core innovation lies in combining interpretable logistic regression with fine-tuned large language models (LLMs). This allows for a robust classification of depressive symptom status using objective wearable-derived behavior, comprehensive clinical features, and relevant environmental exposures. The system aims to provide not just a diagnosis, but also an understanding of why a particular assessment is made, enhancing trust and clinical utility.

Depression affects approximately 280 million individuals worldwide, significantly impairing well-being, social functioning, and physical health. Early detection is a public health imperative, particularly in regions like South Korea with high suicide rates where depression is a major contributing factor. Current clinical assessments like PHQ-9 are crucial but often retrospective and episodic.

Wearable technologies offer continuous, unobtrusive monitoring, providing granular behavioral and physiological data that can enrich mental health assessments. Integrating this with environmental data (e.g., temperature, precipitation) and considering contextual factors like weekday/weekend routines is essential for accurate interpretation and predictive modeling. The challenge lies in developing flexible, interpretable models that can leverage this rich digital phenotyping data.

The study found that 13.0% of participants had elevated depressive symptoms. Individuals with higher depressive symptoms were younger, had higher body weight and resting pulse rate, and were more likely to be unmarried, less educated, unemployed, and living alone. Lifestyle factors like smoking and alcohol consumption were also more prevalent. Wearable data showed that those with higher symptoms had significantly fewer non-daylight steps and lower vigorous physical activity, especially on weekdays.

The OR-based fine-tuned LLM model significantly outperformed traditional nomograms and label-based LLM fine-tuning. It achieved 90.3% accuracy and 94.5% specificity on weekdays, and 88.8% accuracy and 96.4% specificity on weekends. Wearable-derived features provided statistically significant incremental predictive value beyond demographic and clinical variables, with larger gains on weekdays. Inference latency was low (around 3.8 seconds) with compact outputs.

The study acknowledges several limitations, including the PHQ-9 cutoff not being equivalent to a clinician-adjudicated diagnosis, the single-center cohort (Chonnam National University Hospital) with a higher-than-general-population prevalence of depression, and the reliance on GPT-4o for fine-tuning due to infrastructure constraints. The temporal alignment of PHQ-9 (14-day recall) with 7-day actigraphy data means the models provide near-concurrent screening rather than prospective prediction.

Future work should involve prospective external validation in diverse community-based cohorts to confirm generalizability. Evaluating fine-tuning across multiple LLM families and assessing the impact of alternative PHQ-9 thresholds or continuous outcome modeling are also important. Further research should explore alternative feature engineering strategies, longer monitoring durations, and rigorous evaluation of the chatbot interface with clinicians and patients to ensure functional interpretability and usability in real-world clinical workflows.