LLMCARE: Alzheimer’s Detection via Transformer Models Enhanced by LLM-Generated Synthetic Data

Overview

This repository implements LLMCARE, a speech-based NLP pipeline for early detection of Alzheimer’s disease and related dementias (ADRD). The pipeline integrates:

1. Transformer embeddings with handcrafted linguistic features.
2. Synthetic data augmentation using large language models (LLMs).
3. LLMs as classifiers (text-only and multimodal).
4. External generalizability evaluation using the DementiaBank Delaware Corpus.

The repository is organized into five components, each with its own README and runnable code.

---

Components

Component 1 – Transformer & Fusion Screening

• Evaluates ten transformer models (general-purpose + biomedical/clinical).
• Compares frozen, last-layer, and full fine-tuning.
• Incorporates 110 handcrafted linguistic features.
• Develops a fusion classifier combining embeddings + features. 👉 See Component 1 README

---

Component 2 – Synthetic Text Augmentation

• Uses LLMs (e.g., LLaMA-8B/70B, MedAlpaca-7B, Ministral-8B, GPT-4o) to generate label-conditioned synthetic speech transcripts.
• Augmented data boosts F1 scores (e.g., MedAlpaca +2× augmentation improved F1 to 85.7). 👉 See Component 2 README

---

Component 3 – LLMs as Text-Only Classifiers

• Evaluates fine-tuned LLMs as direct classifiers for speech transcripts.
• Demonstrates significant performance gains with fine-tuning. 👉 See Component 3 README

---

Component 4 – Multimodal LLMs

• Benchmarks multimodal LLMs (GPT-4o, Qwen 2.5 Omni, Phi-4) for audio-text classification.

• Includes two subfolders:

  • Qwen → Fine-tuning with LLaMA-Factory.

    • Train with: llamafactory-cli train train.yaml
    • Inference with: test_audio_classification.py

  • Phi4 → Fine-tuning with native scripts.

    • Train with: bash run.sh or python finetune.py
    • Inference with: bash test.sh or python test.py 👉 See Component 4 README

---

Component 5 – External Generalizability Evaluation (DementiaBank Delaware Corpus)

• Tests pipeline generalizability on an MCI-only cohort (n=205).
• Tasks: Cookie Theft, Cinderella recall, procedural discourse.
• Pipeline achieved F1 = 72.8 (AUC = 69.6) with MedAlpaca augmentation.
• Note: This codebase is the same as Component 1 (ADReSSo pipeline). To run, simply update data paths in configs to point to the Delaware dataset.

---