Enabling Few-Shot Alzheimer's Disease Diagnosis on Biomarker Data with Tabular LLMs

Area of Science:

• Biomedical Informatics
• Artificial Intelligence in Healthcare
• Neurodegenerative Disease Research

Background:

• Alzheimer's disease (AD) diagnosis relies on analyzing diverse biomarkers, often in tabular formats.
• Large language models (LLMs) show promise for predicting outcomes from structured biomedical data due to their reasoning and integration capabilities.

Purpose of the Study:

• To adapt a multimodal tabular-specialized LLM (TableGPT2) for Alzheimer's disease diagnosis using structured biomarker data.
• To develop a novel framework, TAP-GPT (Tabular Alzheimer's Prediction GPT), for accurate AD prediction with limited sample sizes.

Main Methods:

• Constructed few-shot tabular prompts using in-context learning from structured biomedical data.
• Fine-tuned TableGPT2 using parameter-efficient qLoRA for binary classification of AD versus cognitively normal (CN).
• Evaluated TAP-GPT against general-purpose LLMs and a tabular foundation model (TFM).

Main Results:

• TAP-GPT demonstrated superior performance in predicting Alzheimer's disease compared to advanced general-purpose LLMs.
• The framework outperformed a specialized tabular foundation model (TFM) in the AD/CN classification task.
• Achieved accurate predictions using structured biomarker data with small sample sizes.

Conclusions:

• TAP-GPT represents the first application of LLMs for prediction tasks using tabular biomarker data in Alzheimer's disease research.
• This framework offers a new avenue for leveraging LLMs in biomedical informatics for disease prediction.
• Paves the way for future LLM-driven multi-agent systems in clinical diagnostics.