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

Abstract

Recent advances in artificial intelligence, particularly large language models (LLMs), show promise for mental health applications, including the automated detection of depressive symptoms from natural language. We fine-tuned a German BERT-based LLM to predict individual Montgomery-Åsberg Depression Rating Scale (MADRS) scores using a regression approach across different symptom items (0-6 severity scale), based on structured clinical interviews with transdiagnostic patients as well as synthetically generated interviews. The fine-tuned model achieved a mean absolute error of 0.7-1.0 across items, with accuracies ranging from 79 to 88%, closely matching clinician ratings. Fine-tuning resulted in a 75% reduction in prediction errors relative to the untrained model. These findings demonstrate the potential of lightweight LLMs to accurately assess depressive symptom severity, offering a scalable tool for clinical decision-making, and monitoring treatment progress, particularly in low-resource settings.