Reading between the Lines: Leveraging Large Language Models for Global Dementia and Depression Assessment from Clinical Interviews

A study of 154 German-speaking subjects finds that open-weights large language models can predict depression severity from clinical interview speech without task-specific training, while dementia assessment benefits from structured feature extraction, according to research posted to arXiv on June 16, 2026 [1]. Dementia and depression are the most prevalent neuropsychiatric disorders in geriatric populations, and their overlapping symptoms complicate differential diagnosis [1]. The study introduces an observer-based Global Depression Scale, aligned with the established Global Deterioration Scale, to enable parallel staging of affective and cognitive symptoms [1]. Researchers compared three large language models — Mistral 3.1, DeepHermes, and Qwen3 — across two settings: zero-shot prediction and LLM-based feature extraction for Support Vector Regression [1]. For depression severity, the models performed best in zero-shot settings, achieving a mean absolute error of 0.60 [1]. Dementia assessment showed stronger results with structured feature extraction, reaching a best MAE of 0.78 and reducing errors by up to 35 percent over zero-shot baselines [1]. The work used both human transcriptions and pause-enriched transcripts, with the latter achieving competitive performance, pointing toward fully automatic screening pipelines [1]. Transcription quality is a known variable in speech-based clinical tools. In molecular biology, transcription factors are proteins that control the rate of genetic information transfer from DNA to messenger RNA by binding to specific DNA sequences, a process fundamental to gene regulation [6]. While unrelated to audio transcription, the precision required in biological transcription underscores the broader principle that fidelity in converting one form of information to another is critical for downstream analysis. The study’s use of pause-enriched transcripts — which preserve temporal features of speech — reflects a parallel emphasis on retaining signal fidelity for machine interpretation [1]. The research contributes to a growing body of work applying machine learning to clinical speech analysis. Prior work in other domains, such as catalysis informatics, has explored whether new datasets complement existing ones for training models, including through transfer learning and joint training across datasets [3]. The dementia and depression study does not employ transfer learning, but its comparison of zero-shot and feature-extraction approaches addresses a similar question of how best to leverage pre-trained models for specialized tasks [1]. The findings suggest that LLMs can extract clinically relevant information from unstructured speech, though the performance gap between depression and dementia prediction indicates that task-specific structuring remains valuable for cognitive assessment [1]. The study’s fully automatic pipeline, combining automatic speech recognition with LLM-based analysis, could reduce reliance on human transcription in screening settings [1].