Title: Assessing Large Language Models’ Knowledge of Alzheimer’s Disease: A Study on Multi-Loss Fine-Tuning and Probing for Detection

Abstract:

Achieving reliable early diagnosis of Alzheimer’s disease (AD) is a significant hurdle, primarily driven by the scarcity of labeled datasets. Although large language models (LLMs) have demonstrated robust transfer learning capabilities across various fields, their adaptation to the medical domain of AD via supervised fine-tuning has received little attention. This study empirically assesses the efficacy of diverse model architectures for text-based AD detection by utilizing three distinct transcript corpora: Pitt, CCC, and ADRC. Furthermore, we examine how task-specific information is embedded within the models' internal representations.

To the best of our knowledge, our fine-tuned BERT and T5 models set a new state-of-the-art benchmark on the Pitt and CCC datasets, while also delivering strong performance on the ADRC corpus. Concurrently, the decoder-only Llama-1B model yields highly competitive outcomes, matching the performance of BERT and T5 across all three datasets, which underscores its potential for AD detection. We perform a thorough evaluation of the Llama-1B backbone, investigating cross-corpus transferability, the influence of clinical transcript markers, and the optimal granularity for input chunk sizes. Additionally, through linear probing, we provide empirical evidence that fine-tuning alters the representations of both linguistic markers and content words in ways that align with AD-related signals.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC