🧬 biology

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

The paper introduces TAP-GPT, a domain-adapted tabular large language model that leverages few-shot learning and structured reasoning to achieve accurate, interpretable, and robust Alzheimer's disease prediction using multimodal biomedical data, outperforming traditional machine learning baselines in scenarios with small or incomplete datasets.

Original authors: Sophie Kearney, Shu Yang, Zixuan Wen, Weimin Lyu, Bojian Hou, Duy Duong-Tran, Tianlong Chen, Jason H. Moore, Marylyn D. Ritchie, Chao Chen, Li Shen

Published 2026-03-19

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Sophie Kearney, Shu Yang, Zixuan Wen, Weimin Lyu, Bojian Hou, Duy Duong-Tran, Tianlong Chen, Jason H. Moore, Marylyn D. Ritchie, Chao Chen, Li Shen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Picture: A New Kind of Medical Detective

Imagine you are trying to diagnose Alzheimer's disease. Traditionally, doctors look at a patient's medical chart—a table full of numbers like brain scan volumes, protein levels in spinal fluid, and genetic markers.

For a long time, computers have tried to read these charts.

Old-school computers (like Random Forests) are like experienced accountants. They are great at crunching numbers and finding patterns, but they can't explain why they made a decision. They just give you a "Yes" or "No."
Deep Learning models are like super-fast calculators, but they often get confused if they don't have enough data to study. If the patient's file is missing a few pages or the group of patients is small, these calculators often fail.

Enter TAP-GPT. Think of TAP-GPT as a brilliant medical student who has read every medical textbook in the world (thanks to being a Large Language Model) but has also been specifically trained to read medical charts.

This paper introduces a new tool called TAP-GPT (Tabular Alzheimer's Prediction GPT). It's designed to look at small, messy medical tables and predict if a patient has Alzheimer's, all while explaining its reasoning in plain English.

How It Works: The "Study Buddy" Analogy

1. The Problem: Small, Messy Data

In medical research, you often don't have millions of patient records. You might only have a few dozen. Also, patients often miss appointments, so their files have holes (missing data).

The Old Way: If you try to teach a computer with only 10 examples, it usually panics and guesses randomly.
The TAP-GPT Way: TAP-GPT uses a trick called "Few-Shot Learning." Imagine you are taking a test. Instead of studying for years, you are given three example questions with answers right before the test. TAP-GPT is incredibly good at looking at those few examples, understanding the pattern, and applying it to the new patient.

2. The Secret Sauce: "Tabular" vs. "Text"

Most AI models are trained on text (like books or tweets). If you feed them a spreadsheet, they often get confused because spreadsheets don't have sentences; they have rows and columns.

The Innovation: The researchers didn't just turn the spreadsheet into text. They used a special version of an AI called TableGPT2 that understands the structure of a table.
The Analogy: Imagine trying to describe a chessboard.
- Serialized (Text): "The white king is on E1, the black pawn is on A7..." (Hard to visualize).
- Tabular (TAP-GPT): The AI sees the actual grid. It understands that the King is next to the Rook, not just that they are mentioned in the same sentence. TAP-GPT keeps the "grid" structure intact, making it much smarter at reading medical charts.

3. The "Explainable" Superpower

Old AI models are "black boxes." You ask, "Does this patient have Alzheimer's?" and it says "Yes." You have no idea why.

TAP-GPT's Approach: It acts like a detective writing a case file. When it predicts "Yes," it also writes a note: "I think this patient has Alzheimer's because their hippocampus (memory center) is shrunken, and their amyloid protein levels are high, similar to the three patients I just looked at."
Why this matters: Doctors can trust the AI more because they can see the logic. It's not magic; it's reasoning.

What Did They Test?

The researchers tested TAP-GPT on four different types of "medical charts" from the Alzheimer's Disease Neuroimaging Initiative (ADNI):

QT-PAD: A mix of blood tests, genetic markers, and brain scans.
Amyloid PET: Scans showing sticky proteins in the brain.
Tau PET: Scans showing tangles in the brain.
MRI: Standard brain scans showing brain shrinkage.

The Results:

Beating the Old Guard: TAP-GPT beat the traditional "accountant" computers (like Random Forests) and even the specialized "tabular AI" (TabPFN) when the data was small and messy.
Handling Missing Pages: In the real world, patients often miss tests. If a patient's file is missing 30% of the data, most computers crash or guess wildly. TAP-GPT, however, is like a detective who can solve a case even if half the evidence is missing. It just looks at what is there and uses its general medical knowledge to fill in the gaps.
The "Self-Reflection" Test: The researchers asked the AI, "Are you sure?" and let it rethink its answer.
- Other AIs: Often got confused and changed their minds to the wrong answer.
- TAP-GPT: Stayed steady. It was like a confident doctor who, after double-checking their notes, stuck with the correct diagnosis.

The "Feature Selection" Trick

One challenge was that some brain scans have 72 different measurements. That's like trying to read a book where every page has 72 different fonts. It's too much information for the AI to process quickly.

The researchers taught TAP-GPT to be a curator. Before reading the whole chart, it uses a filter (called LASSO) to pick the top 16 most important clues.

Analogy: Imagine a detective arriving at a crime scene with 100 pieces of evidence. Instead of looking at everything, they immediately grab the gun, the fingerprint, and the witness statement. TAP-GPT learned to grab the "gun and fingerprints" of Alzheimer's (specific brain regions) and ignore the noise. This made it much faster and more accurate.

Why This Matters for the Future

This paper isn't just about predicting Alzheimer's; it's about how we use AI in medicine.

Trust: Because TAP-GPT explains its reasoning, doctors can actually use it as a partner, not just a calculator.
Teamwork (Multi-Agent Systems): The authors imagine a future where different AI agents talk to each other. One AI agent looks at the MRI, another looks at the blood test, and they "debate" the diagnosis. Because TAP-GPT is stable and can explain itself, it's perfect for being part of this team.
Real-World Ready: It works even when data is messy or incomplete, which is exactly how real hospitals operate.

The Bottom Line

TAP-GPT is a medical AI that doesn't just guess; it reasons. It combines the pattern-finding power of a computer with the logical explanation skills of a human doctor. It can look at a small, messy medical chart, spot the signs of Alzheimer's, and tell you exactly why it thinks so, making it a powerful new tool for early detection and better patient care.

1. Problem Statement

Accurate diagnosis of Alzheimer's Disease (AD) relies on heterogeneous, multimodal biomedical data (e.g., MRI, PET, CSF biomarkers, genetics). However, clinical datasets often suffer from three critical challenges:

Small Sample Sizes: Biomedical cohorts are often small relative to the number of features (high dimensionality), causing deep learning models to overfit or fail to outperform classical methods.
Missing Data: Real-world clinical data frequently contain missing values, requiring complex imputation strategies that can introduce bias.
Lack of Interpretability: Traditional machine learning models (e.g., Random Forest, XGBoost) and deep neural networks act as "black boxes," offering limited insight into why a prediction was made, which is crucial for clinical trust and decision-making.

While Large Language Models (LLMs) offer few-shot generalization and reasoning capabilities, standard LLMs struggle with tabular data due to the lack of natural ordering in columns and token inefficiency when serializing tables. Conversely, Tabular Foundation Models (TFMs) like TabPFN excel at small datasets but lack natural language reasoning and interpretability.

2. Methodology: TAP-GPT Framework

The authors propose TAP-GPT (Tabular Alzheimer's Prediction GPT), a domain-adapted tabular LLM framework designed for few-shot AD classification.

Architecture: TAP-GPT is built upon TableGPT2 (a 7B parameter model based on Qwen2.5), which features a dedicated semantic table encoder and an LLM decoder.
- Encoder: Fixed during training to preserve learned structural representations of tabular data.
- Decoder: Fine-tuned using QLoRA (Quantized Low-Rank Adaptation) to specialize in binary AD/CN classification without full weight retraining.
Input Strategy (Few-Shot In-Context Learning):
- The task is formulated as predicting a target patient's label based on a prompt containing $k$ labeled in-context examples (ICL) and the target patient.
- Prompt Formats: The study evaluates four formats:
  1. Zero-Shot Tabular: Single row.
  2. Zero-Shot Serialized: Natural language description.
  3. Few-Shot Tabular: $k$ examples + target in table format.
  4. Few-Shot Serialized: $k$ examples + target in text format.
Feature Selection: To handle high-dimensional imaging data (up to 74 regions), the authors employ LASSO and LLM-guided ranking to select optimal feature subsets ( $p$ ) before prompt construction, mitigating context-length constraints.
Interpretability: The model is prompted to output a structured JSON object containing the prediction, a confidence score, and a step-by-step reasoning explanation (Chain-of-Thought).

3. Key Contributions

First Systematic Application: This is the first study to apply a tabular-specialized LLM to multimodal AD biomarker prediction, bridging the gap between structured tabular learning and semantic reasoning.
Domain Adaptation: Successfully adapted TableGPT2 for AD diagnosis using parameter-efficient fine-tuning on AT[N] (Amyloid, Tau, Neurodegeneration) biomarkers and imaging data.
Robustness to Missingness: Demonstrated that TAP-GPT can handle missing data natively within the prompt without explicit imputation, maintaining stability under both simulated and real-world missingness.
Interpretability & Stability: Showed that TAP-GPT produces clinically aligned reasoning and exhibits greater stability under self-reflection (iterative reasoning) compared to general-purpose LLMs, making it suitable for multi-agent diagnostic systems.
Comprehensive Benchmarking: Evaluated across four datasets (QT-PAD, Amyloid PET, Tau PET, Structural MRI) against classical ML, TabPFN, and state-of-the-art LLMs (GPT-4.1-mini, Qwen3).

4. Key Results

Performance on Biomarker Data (QT-PAD):
- TAP-GPT significantly outperformed traditional ML (Logistic Regression, XGBoost, Random Forest) and vanilla LLMs in the few-shot tabular setting.
- Interpretable TAP-GPT (with CoT reasoning) achieved the highest zero-shot performance (Mean F1 $\approx$ 0.89).
- Optimal in-context examples ( $k$ ) were found to be 8.
Performance on Imaging Data (PET/MRI):
- TAP-GPT consistently outperformed traditional ML and TabPFN in low-shot settings (4-shot).
- Feature Selection is Critical: Using all 72 regional features degraded performance significantly. Selecting the top 16 features via LASSO restored high performance.
- TAP-GPT performed competitively against GPT-4.1-mini, often surpassing it in Tau PET and MRI modalities, despite GPT-4.1-mini being a more powerful general-purpose model.
Robustness to Missingness:
- TAP-GPT maintained high performance up to 40% simulated missingness.
- In real-world missingness scenarios (naturally occurring gaps in ADNI data), TAP-GPT showed stable performance across varying levels of target missingness, outperforming baselines that required imputation.
Self-Reflection Stability:
- When prompted to review and revise their own predictions, TAP-GPT maintained performance, whereas GPT-4.1-mini suffered significant performance drops. This suggests TAP-GPT is more robust for iterative multi-agent workflows.

5. Significance and Future Implications

Paradigm Shift: The paper demonstrates that pretrained tabular LLMs can effectively replace or augment classical machine learning in low-data clinical settings, offering a new paradigm that combines statistical learning with semantic reasoning.
Clinical Utility: The ability to generate structured, modality-aware explanations (e.g., linking hippocampal atrophy to AD diagnosis) provides a level of transparency essential for clinical adoption.
Multi-Agent Systems: The stability of TAP-GPT under self-reflection and its ability to handle missing data make it a strong candidate for multi-agent diagnostic systems, where different AI agents (specialized in MRI, genetics, or PET) can collaborate, critique, and refine a diagnosis without destabilizing the final output.
Open Science: The source code and models are publicly available, facilitating further research into tabular LLMs for healthcare.

In conclusion, TAP-GPT establishes that domain-adapted tabular LLMs are a viable, robust, and interpretable solution for Alzheimer's disease prediction, effectively addressing the challenges of small sample sizes, missing data, and the need for explainable AI in precision health.