Large Language Model Adaptation Strategies in Speech-Based Cognitive Screening: Systematic Evaluation.

Area of Science:

• Natural Language Processing
• Artificial Intelligence in Healthcare
• Speech Analysis

Background:

• Over 50% of US adults with Alzheimer disease and related dementias (ADRD) are undiagnosed.
• Speech-based screening offers a scalable solution, but the effectiveness of large language model (LLM) adaptation strategies requires further investigation.

Purpose of the Study:

• To compare various LLM adaptation strategies for detecting cognitive impairment using speech data.
• To evaluate both text-only and multimodal LLM approaches on DementiaBank datasets.

Main Methods:

• Analysis of audio-recorded speech from 237 participants (ADRD vs. cognitive normal) in the ADReSSo dataset.
• Evaluation of nine text-only LLMs and three multimodal models using adaptation strategies: in-context learning (ICL), reasoning-augmented prompting, and parameter-efficient fine-tuning.
• Assessment of strategy generalizability on the DementiaBank Delaware dataset (mild cognitive impairment vs. cognitive normal).

Main Results:

• Prototype demonstrations in ICL yielded the highest performance (F1-score up to 0.81) on the ADReSSo dataset.
• Token-level fine-tuning achieved the highest scores across models (e.g., LLaMA 3B: F1=0.83, AUC=0.91).
• Reasoning-augmented prompting benefited smaller models, while multimodal models did not outperform top text-only systems.

Conclusions:

• Detection accuracy depends on demonstration selection, reasoning design, and tuning methods.
• Token-level fine-tuning is generally most effective for speech-based ADRD and mild cognitive impairment screening.
• Adapted open-weight models show potential to match or surpass commercial LLMs; multimodal models may need further refinement.