Tabular LLMs for Interpretable Few-Shot Alzheimer's Disease Prexdiction with Multimodal Biomedical Data.

Area of Science:

• Artificial Intelligence in Medicine
• Biomedical Informatics
• Computational Neuroscience

Background:

• Accurate Alzheimer's disease (AD) diagnosis relies on tabular biomarker data, which is often small and incomplete.
• Traditional deep learning models struggle with such data, frequently underperforming classical methods.
• Pretrained large language models (LLMs) offer potential for clinical prediction due to their generalization, reasoning, and interpretability.

Purpose of the Study:

• To introduce TAP-GPT, a domain-adapted tabular LLM framework for few-shot Alzheimer's disease classification.
• To evaluate TAP-GPT's performance on multimodal and unimodal AD biomarker datasets.
• To demonstrate the efficacy of tabular LLMs in clinical prediction tasks with limited and incomplete data.

Main Methods:

• Developed TAP-GPT, a framework fine-tuned on TableGPT2 using tabular prompts for AD classification.
• Evaluated TAP-GPT on four ADNI-derived datasets including QT-PAD biomarkers and neuroimaging data (MRI, amyloid PET, tau PET).
• Assessed performance in few-shot settings, comparing against backbone models, traditional machine learning, and general-purpose LLMs, while testing robustness to missing data.

Main Results:

• TAP-GPT improved upon backbone models and outperformed traditional machine learning baselines in few-shot AD classification.
• The framework demonstrated competitive performance with state-of-the-art general-purpose LLMs.
• TAP-GPT maintained stable performance with missing data without imputation and produced interpretable, biology-aligned reasoning.

Conclusions:

• TAP-GPT represents the first systematic application of a tabular-specialized LLM for multimodal biomarker-based AD prediction.
• Domain-adapted tabular LLMs can effectively handle structured clinical prediction tasks, even with incomplete data.
• TAP-GPT lays the groundwork for multi-agent clinical decision-support systems driven by tabular LLMs.