Research map / Large language models / Language, safety and interpretability
DECT: Harnessing LLM-assisted Fine-Grained Linguistic Knowledge and Label-Switched and Label-Preserved Data Generation for Diagnosis of Alzheimer’s Disease
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DECT, an LLM-assisted pipeline that extracts cognitive-linguistic atoms and linguistic markers from ADReSSo transcripts and generates label-switched/label-preserved synthetic transcripts to fine-tune a BioBERT-based Alzheimer's detector.
Problem
Speech-based Alzheimer's disease detection from noisy patient-interviewer transcripts is hard due to scarce labeled data, speech variability, and irrelevant content.
Approach
Use frozen GPT LLMs to distill CL atoms and linguistic markers from transcripts, then generate label-switched and label-preserved synthetic transcripts with injected AD/non-AD linguistic features, and fuse atom and marker embeddings (AMR) with a text encoder and MLP trained jointly on real and synthetic data.
Key results
- DECT with GPT-4o, AMR input and LSLP generation reaches 90.48% accuracy and 88.32% F1 on ADReSSo, versus 78.57% for BERT-Base and 85.71% for RoBERTa-Large.
- Ablation with GPT-4o: the baseline BioBERT gets 80.95% accuracy; adding atoms and LSLP gives 89.29%; adding markers too gives 90.48%.
- Zero-shot-style LLM baselines perform poorly, with GPT-4o at 75.30% accuracy and open LLMs between 40.97% and 55.22%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CL atom extraction to the component-free model (ablation) | DECT · ADReSSo | accuracy: 80.95\(\pm 2.71\) -> 82.38\(\pm 2.61\) | Helped |
| Add linguistic marker identification to the component-free model (ablation) | DECT · ADReSSo | accuracy: 80.95\(\pm 2.71\) -> 66.67\(\pm 2.38\) | Hurt |
| Add LSLP generation without fine-grained representations (ablation) | DECT · ADReSSo | accuracy: 80.95\(\pm 2.71\) -> 76.66\(\pm 2.61\) | Hurt |
| Add CL atoms and LSLP generation (ablation) | DECT · ADReSSo | accuracy: 80.95\(\pm 2.71\) -> 89.29\(\pm 1.31\) | Helped |
| Add CL atoms, linguistic markers, and LSLP generation (ablation) | DECT · ADReSSo | accuracy: 80.95\(\pm 2.71\) -> 90.48\(\pm 2.38\) | Helped |
| Use LSLP instead of Mimic generation with transcript input (ablation) | GPT4o-based DECT · ADReSSo | accuracy: 83.33 -> 84.12 | Helped |
| Use LSLP instead of Mimic generation with AMR input (ablation) | GPT4o-based DECT · ADReSSo | accuracy: 88.88 -> 90.48 | Helped |
| Use AMR instead of transcript input with GPT4o and LSLP (ablation) | DECT · ADReSSo | accuracy: 84.12 -> 90.48 | Helped |
| Use GPT4o instead of GPT4 as the DECT LLM | DECT · ADReSSo | accuracy: 85.31\(\pm 2.52\) -> 90.48\(\pm 2.38\) | Helped |
Surprising findings
- Generating data alone without linguistic feature injection lowered accuracy (76.66% vs 80.95% baseline), per the ablation.
- Standalone LLMs performed worse than fine-tuned BERT-style encoders, with several near chance level.
Limitations
- Small evaluation set (166 samples) and only one dataset, ADReSSo.
- Synthetic data distribution is not controlled, risking spurious correlations and overfitting, as the authors note.
- Heavy reliance on closed-source GPT models for extraction and generation.
Experimental setup
Datasets: ADReSSo DementiaBank
Benchmarks: ADReSSo
Metrics: accuracy F1
Baselines: BERT-Base BERT-Large FLAN-T5-Large RoBERTa-Base RoBERTa-Large BioBERT MedBERT Qwen1.5-7B-Chat Qwen2-7B-Instruct ChatGLM3-6B GLM-4-9B-Chat Yi-1.5-9B-Chat-16K
Builds on: BioBERT Whisper factual atoms (Stacey et al.)
Closest papers by meaning
- Beyond Plain Demos: A Demo-Centric Anchoring Paradigm for In-Context Learning in Alzheimer’s Disease Detection (AAAI 2026 · similarity 0.88)
- EMAD: Evidence-Centric Grounded Multimodal Diagnosis for Alzheimer's Disease (CVPR 2026 · similarity 0.86)
- Language Model Augmented Semi-Supervised Statistical Inference (ICML 2026 · similarity 0.86)
- The Brain's Bitter Lesson: Scaling Speech Decoding With Self-Supervised Learning (ICML 2025 · similarity 0.83)
- LibriBrain100: One Hundred Hours of Broad and Deep MEG Data for Neural Speech Decoding at Scale (NeurIPS 2026 · similarity 0.83)
- Knowledge Is Power: Harnessing Large Language Models for Enhanced Cognitive Diagnosis (AAAI 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.