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

TL;DR

This work tackles the challenge of diagnosing Alzheimer's disease from tabular biomarker data in data-scarce regimes. It introduces TAP-GPT, a domain-adapted tabular LLM built by finetuning TableGPT2 with qLoRA on the QT-PAD biomarker dataset, enabling few-shot in-context learning for AD vs CN classification. Results show TAP-GPT achieving top performance in few-shot tabular prompts and competitive results in low-shot settings, while also offering natural-language explanations for predictions, addressing the critical need for interpretability in clinical deployments. The study demonstrates the viability of combining tabular encoders with language-model reasoning to leverage prior biomedical knowledge, paving the way for multi-agent health informatics frameworks and broader applications to tabular biomedical data.

Abstract

Early and accurate diagnosis of Alzheimer's disease (AD), a complex neurodegenerative disorder, requires analysis of heterogeneous biomarkers (e.g., neuroimaging, genetic risk factors, cognitive tests, and cerebrospinal fluid proteins) typically represented in a tabular format. With flexible few-shot reasoning, multimodal integration, and natural-language-based interpretability, large language models (LLMs) offer unprecedented opportunities for prediction with structured biomedical data. We propose a novel framework called TAP-GPT, Tabular Alzheimer's Prediction GPT, that adapts TableGPT2, a multimodal tabular-specialized LLM originally developed for business intelligence tasks, for AD diagnosis using structured biomarker data with small sample sizes. Our approach constructs few-shot tabular prompts using in-context learning examples from structured biomedical data and finetunes TableGPT2 using the parameter-efficient qLoRA adaption for a clinical binary classification task of AD or cognitively normal (CN). The TAP-GPT framework harnesses the powerful tabular understanding ability of TableGPT2 and the encoded prior knowledge of LLMs to outperform more advanced general-purpose LLMs and a tabular foundation model (TFM) developed for prediction tasks. To our knowledge, this is the first application of LLMs to the prediction task using tabular biomarker data, paving the way for future LLM-driven multi-agent frameworks in biomedical informatics.

Figures (5)

• Figure 1: Overview of the TAP-GPT framework. As illustrated in the QT-PAD Biomarker Tabular Data, we split the AD and control subjects into pools for ICL (green), training (blue), and testing (yellow), from which we construct tables used in the downstream tasks (finetuning and inference). CN: cognitively normal. AD: Alzheimer's disease. ICL: in context learning. LLM: large language model. TFM: tabular foundation model.
• Figure 2: Prompt formats and models used in our experiments. Prompts provide subject data to LLMs, while TabPFN takes tabular CSV input. Serialized formats (blue) and tabular formats (green) are consistent across figures. Fire icons mark finetuned parameters; snowflakes mark frozen parameters.
• Figure 3: Mean F1 for AD diagnosis across models in zero-shot and few-shot ($k=8$) contexts. LLMs use tabular (green) and serialized (blue) prompts with error bars for standard deviation; TabPFN and traditional ML (yellow) operate directly on structured data. TAP-GPT achieves the best performance in the tabular few-shot setting, surpassing TabPFN, general LLMs, and ML baselines.
• Figure 4: k ablation analysis across TableGPT2, TabPFN, and TAP-GPT. TabPFN performance steadily improved with larger k, TableGPT2 improved up to $k=6$ and remained stable thereafter, and TAP-GPT peaked in the middle range.
• Figure 5: Evaluation of TableGPT2 finetuned on zero-shot prompts used on few-shot prompts. Performance drops significantly, indicating that the finetuned model does not generalize well across prompt formats.