Introduction
Early detection of Alzheimer’s Disease (AD) is critical for effective intervention, and language impairment often serves as one of the earliest clinical indicators. However, developing automated diagnostic tools based on speech transcripts faces two primary challenges: noisy data and data scarcity. Real-world patient-interviewer dialogues are often cluttered with interviewer interruptions and irrelevant "filler" talk, while high-quality, labeled AD speech datasets remain small compared to standard NLP benchmarks.
To address these issues, researchers have introduced DECT (Domain-specific LLM-assisted approach), a framework designed to transform noisy, scarce transcripts into high-quality diagnostic data by leveraging the feature extraction and data augmentation capabilities of Large Language Models (LLMs).
The DECT Framework
The DECT pipeline consists of four main stages designed to "supercharge" a standard text encoder, such as BioBERT, for clinical classification.
- Distilling Cognitive-Linguistic (CL) Atoms: The LLM filters out interviewer questions and conversational noise to extract concise, factual statements (e.g., "The boy is taking cookies").
- Extracting Linguistic Markers: Simultaneously, the LLM identifies stylistic speech patterns indicative of cognitive decline, such as repetition, word-finding difficulties, or sudden topic switches.
- LSLP Data Generation: To mitigate data scarcity, the framework generates Label-Switched and Label-Preserved (LSLP) data. By taking factual "atoms" from healthy individuals and rewriting them with AD-specific linguistic markers (and vice versa), the model creates a diverse synthetic dataset that clarifies the decision boundary between healthy and impaired speech.
- Atom-Marker Representation (AMR): Finally, the factual content (Atoms) and stylistic patterns (Markers) are fused into a single embedding to fine-tune a task-specific diagnostic model.
Experimental Results
The study utilized the ADReSSo Dataset, a standard clinical assessment benchmark where participants describe the "Cookie Theft" picture. The dataset contains 166 balanced samples (86 AD, 80 Normal Control).
The DECT framework demonstrated significant performance improvements over both traditional NLP baselines and standalone LLMs. Using GPT-4o as the assistant, DECT achieved a state-of-the-art 90.48% accuracy and an 88.32% F1-score.
As shown in the results table, DECT represents an 11% improvement over the best non-LLM baseline (RoBERTa-Large). Notably, while standalone LLMs like GPT-4o perform reasonably well (75.30% accuracy), they are significantly outperformed by the specialized DECT framework, which uses the LLM as a feature extractor rather than a direct classifier.
Discussion and Conclusion
A key insight from this research is that LLMs are most effective in clinical settings as "assistants" for data cleaning and augmentation. By separating what a patient says (factual atoms) from how they say it (linguistic markers), DECT provides a more robust representation for diagnosis. The LSLP strategy further strengthens the model by generating challenging synthetic samples near the decision boundary.
Despite these gains, the study is limited by a relatively small sample size and a cross-sectional focus. Future research should explore longitudinal analysis to track linguistic changes over time and refine LLM generation to ensure clinical accuracy in synthetic data.
Reference: Mo, T., Lam, J. C. K., Li, V. O. K., & Cheung, L. Y. L. (2025). DECT: Harnessing LLM-assisted Fine-Grained Linguistic Knowledge and Label-Switched and Label-Preserved Data Generation for Diagnosis of Alzheimer’s Disease. The Thirty-Ninth AAAI Conference on Artificial Intelligence (AAAI-25).
Summary updated