Related Experiment Video
Updated: Apr 20, 2026
12:18
A Machine Learning Approach to Design an Efficient Selective Screening of Mild Cognitive Impairment
Published on: January 11, 2020
8.2K
Interpretable depressive symptoms screening via statistical reasoning-augmented large language models using wearable
Seokjin Kong1, Yihyun Kim1, Inyong Jeong1
1Department of Biomedical Informatics, Korea University College of Medicine, 161 Jeongneung-ro, Seongbuk-gu, Seoul, 02708, Republic of Korea.
Scientific Reports
|April 18, 2026
Summary
A novel hybrid system using large language models (LLMs) and wearable data accurately detects depressive symptoms. This approach enhances mental healthcare by integrating behavioral metrics with clinical information for early assessment.
Area of Science:
- Digital Health
- Artificial Intelligence in Medicine
- Mental Health Technology
Background:
- Early detection of depressive symptoms is crucial for effective mental healthcare.
- Scalable and personalized approaches are needed for mental health assessment.
- Wearable technology offers potential for continuous, objective behavioral monitoring.
Purpose of the Study:
- To develop a hybrid clinical decision support system (CDSS) for classifying depressive symptom status.
- To fine-tune a GPT-based large language model (LLM) using odds ratio (OR)-derived statistical reasoning.
- To evaluate the diagnostic accuracy, interpretability, and efficiency of the OR-based LLM compared to traditional models and label-supervised LLMs.
Main Methods:
- Utilized data from 2437 adults, including wearable-derived behavioral metrics (steps, physical activity), clinical features, and environmental exposures.
- Employed multivariable logistic regression to identify significant predictors and their odds ratios (ORs).
- Fine-tuned a GPT-based LLM using two strategies: label-based and OR-based (embedding statistical reasoning).
- Compared the hybrid OR-based LLM against nomogram and label-based LLM approaches.
Main Results:
- Wearable-derived behavioral features provided significant incremental predictive value beyond demographic and clinical variables.
- The OR-based fine-tuned LLM framework demonstrated improved balanced accuracy, reaching 78.7% on weekdays and 68.0% on weekends when combining all features.
- The combined OR-based fine-tuned model achieved high accuracy (up to 90.3%) and specificity (up to 96.4%), outperforming other models.
- Incremental predictive value of wearable data was greater on weekdays than weekends.
Conclusions:
- A hybrid CDSS integrating OR-anchored LLM fine-tuning with wearable data shows promise for accurate depressive symptom classification.
- This approach enhances classification performance while maintaining predictor traceability.
- Wearable behavioral data significantly contributes to predicting depressive symptoms, offering a valuable addition to traditional clinical assessments.
- Further external validation in diverse populations is necessary before clinical implementation.
