| Abstract | Alzheimer’s disease (AD) is a progressive neurodegenerative disorder, where subtle linguistic changes can emerge years before clinical symptoms. However, current large language models (LLMs) struggle to capture these weak signals due to their coherence bias, limiting their effectiveness for early AD detection. To address this challenge, we construct a new dataset, DementiaNet_Text, by transcribing and annotating patient interview recordings, and propose a Symptom-Informed Sequential Enhancement (SISE) pipeline. The pipeline first detects fine-grained linguistic anomalies, then disentangles them into structured symptom vectors, and finally integrates them into the Llama-3 embedding space to guide the model toward clinically relevant cues. Our goal is to improve classification of preclinical AD stages. Experiments show that the proposed Llama3-SISE achieves state-of-the-art results, with F1 gains of +4.8 in the Mid preclinical stage and +5.1 in the Late preclinical stage over the Meta-Llama-3-8B baseline, while maintaining competitive performance in Early and Health categories. In summary, our work demonstrates that explicitly modeling subtle linguistic anomalies provides a robust pathway for effective early AD screening. |
|---|
