What Do LLMs Know About Alzheimer's Disease? Multi-loss Fine-Tuning and Probing for AD Detection

Researchers demonstrate that fine-tuned large language models, particularly BERT, T5, and Llama-1B, achieve state-of-the-art performance in detecting Alzheimer's disease from speech transcripts across multiple datasets. The study reveals how these models encode disease-related linguistic signals through fine-tuning, advancing the potential for early AD diagnosis through text analysis.

This research addresses a critical healthcare challenge by leveraging advances in natural language processing for neurological disease detection. Alzheimer's disease diagnosis traditionally requires expensive biomarkers and cognitive testing, making early detection difficult in resource-limited settings. The authors demonstrate that pretrained language models can be adapted effectively to identify AD-related speech patterns, offering a scalable, non-invasive screening approach.

The study's strength lies in its rigorous evaluation across three independent transcript corpora from different sources, validating generalization beyond single datasets. The competitive performance of Llama-1B alongside larger models like BERT and T5 suggests resource-efficient alternatives exist for deployment in clinical or mobile environments. Linear probing analysis reveals mechanistic insights—fine-tuning doesn't simply memorize patterns but systematically reorganizes internal representations to emphasize AD-relevant linguistic markers, indicating the models capture genuine disease signals rather than artifacts.

For the healthcare and AI sectors, this work bridges academic research and clinical application. Early AD detection from voice samples could enable screening at scale through smartphones or telehealth platforms, democratizing access to neurological assessment. The cross-corpus transferability analysis provides roadmaps for adapting similar approaches to other neurodegenerative diseases or languages.

Future developments should focus on prospective validation with real patient populations, establishing clinical decision thresholds, and integrating these models into diagnostic pipelines. Regulatory pathways for AI-assisted diagnostics remain evolving, and establishing safety standards will be essential before deployment in clinical settings.

• →Fine-tuned BERT and T5 models achieve state-of-the-art Alzheimer's detection on standard benchmark datasets (Pitt and CCC).
• →Llama-1B delivers competitive AD detection performance with lower computational requirements than larger language models.
• →Linear probing analysis proves fine-tuning reorganizes token representations to encode disease-specific linguistic patterns, not memorized artifacts.
• →Cross-corpus evaluation across three heterogeneous datasets demonstrates robust generalization of the approach.
• →Text-based AD detection from transcripts offers a scalable, non-invasive alternative to traditional biomarker-dependent diagnosis.