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

Problem
This study addresses the challenge of differential diagnosis between dementia and depression, two prevalent neuropsychiatric disorders in geriatric populations. The overlapping symptoms complicate accurate assessments, necessitating improved methodologies for evaluating these conditions. The authors present a preprint work that leverages open-weights Large Language Models (LLMs) to predict severity from clinical interviews, filling a gap in the literature regarding automated assessment tools in this domain.

Method
The authors introduce an observer-based Global Depression Scale (GDS-D) that aligns with the established Global Deterioration Scale (GDS), facilitating simultaneous evaluation of cognitive and affective symptoms. They compare three LLMs: Mistral 3.1, DeepHermes, and Qwen3, in two experimental settings: (1) zero-shot prediction directly from transcripts and (2) LLM-based feature extraction for Support Vector Regression (SVR). The dataset comprises speech samples from 154 German-speaking subjects, with both human and pause-enriched transcripts utilized. The models were trained and evaluated on predicting depression and dementia severity, with performance metrics including Mean Absolute Error (MAE).

Results
In zero-shot settings, the best-performing model achieved a MAE of 0.60 for depression severity prediction. For dementia assessment, the structured feature extraction approach significantly improved performance, yielding a best MAE of 0.78, which represents a reduction in error by up to 35% compared to zero-shot baselines. Notably, pause-enriched transcripts demonstrated competitive performance relative to human transcriptions, indicating their potential for use in fully automated screening pipelines.

Limitations
The authors acknowledge that the study is limited by its reliance on a relatively small and homogeneous sample of German-speaking subjects, which may affect the generalizability of the findings. Additionally, the performance of the LLMs may vary with different languages or dialects, and the zero-shot approach may not capture the full complexity of clinical assessments. The authors also note that while the results are promising, further validation in larger, more diverse populations is necessary to establish robustness.

Why it matters
This research has significant implications for the development of automated tools for neuropsychiatric assessment, potentially streamlining the diagnostic process for dementia and depression. The introduction of the GDS-D and the demonstrated efficacy of LLMs in this context could pave the way for more accessible and efficient screening methods in clinical settings. The findings contribute to the growing body of literature on the application of LLMs in healthcare, as discussed in related works on automated diagnostic systems, and are available on arXiv.