Using a fine-tuned large language model for symptom-based depression evaluation

Area of Science:

• Artificial Intelligence in Mental Health
• Computational Psychiatry
• Natural Language Processing for Clinical Assessment

Background:

• Large language models (LLMs) show potential for mental health applications.
• Automated detection of depressive symptoms from natural language is an emerging area.
• Accurate assessment of depression severity is crucial for effective treatment.

Purpose of the Study:

• To fine-tune a German BERT-based LLM for predicting Montgomery-Åsberg Depression Rating Scale (MADRS) scores.
• To evaluate the model's accuracy in assessing depressive symptom severity using regression analysis.
• To explore the utility of LLMs in clinical decision-making and treatment monitoring.

Main Methods:

• Fine-tuning a German BERT-based LLM on structured clinical interviews and synthetic data.
• Utilizing a regression approach to predict individual MADRS scores across symptom items (0-6 severity scale).
• Evaluating model performance using mean absolute error and accuracy metrics.

Main Results:

• The fine-tuned LLM achieved a mean absolute error of 0.7-1.0 across symptom items.
• Prediction accuracies ranged from 79% to 88%, closely matching clinician ratings.
• Fine-tuning reduced prediction errors by 75% compared to the untrained model.

Conclusions:

• Lightweight LLMs can accurately assess depressive symptom severity from natural language.
• This technology offers a scalable tool for clinical decision-making and treatment progress monitoring.
• The model shows particular promise for low-resource mental health settings.