PaperlayerView exact revision on arXiv

Source-linked AI summary

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

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

arXiv:2603.17191v1cs.CLcs.LGq-bio.QM

TL;DR

Alzheimer’s prediction from small, incomplete biomedical tables challenges conventional modeling approaches. The paper introduces TAP-GPT, a TableGPT2-based tabular LLM fine-tuned for few-shot AD classification across multimodal ADNI data, and reports consistent gains over backbone and classical baselines, with robustness to missingness and interpretable reasoning.

  • Problem

    Biomedical Alzheimer’s datasets are often small, high-dimensional, multimodal, and incomplete, while conventional deep learning frequently fails to outperform classical methods.

  • Method

    TAP-GPT adapts TableGPT2 for few-shot AD-versus-CN classification from structured AT[N] biomarker and imaging tables using tabular prompts and parameter-efficient fine-tuning.

  • Results

    Across multimodal and unimodal datasets, TAP-GPT improved upon its backbone models, outperformed classical machine-learning baselines, and remained competitive with general-purpose LLMs and TabPFN.

  • Takeaways & Limitations

    TAP-GPT supports structured clinical prediction with interpretable multimodal reasoning, stable missingness handling without imputation, and self-reflection stability.

Abstract

from arXiv · show

Accurate diagnosis of Alzheimer's disease (AD) requires handling tabular biomarker data, yet such data are often small and incomplete, where deep learning models frequently fail to outperform classical methods. Pretrained large language models (LLMs) offer few-shot generalization, structured reasoning, and interpretable outputs, providing a powerful paradigm shift for clinical prediction. We propose TAP-GPT Tabular Alzheimer's Prediction GPT, a domain-adapted tabular LLM framework built on TableGPT2 and fine-tuned for few-shot AD classification using tabular prompts rather than plain texts. We evaluate TAP-GPT across four ADNI-derived datasets, including QT-PAD biomarkers and region-level structural MRI, amyloid PET, and tau PET for binary AD classification. Across multimodal and unimodal settings, TAP-GPT improves upon its backbone models and outperforms traditional machine learning baselines in the few-shot setting while remaining competitive with state-of-the-art general-purpose LLMs. We show that feature selection mitigates degradation in high-dimensional inputs and that TAP-GPT maintains stable performance under simulated and real-world missingness without imputation. Additionally, TAP-GPT produces structured, modality-aware reasoning aligned with established AD biology and shows greater stability under self-reflection, supporting its use in iterative multi-agent systems. To our knowledge, this is the first systematic application of a tabular-specialized LLM to multimodal biomarker-based AD prediction, demonstrating that such pretrained models can effectively address structured clinical prediction tasks and laying the foundation for tabular LLM-driven multi-agent clinical decision-support systems. The source code is publicly available on GitHub: https://github.com/sophie-kearney/TAP-GPT.

I. INTRODUCTION

TAP-GPT adapts a tabular foundation model for few-shot Alzheimer’s disease classification from structured biomedical data. The framework uses split-specific in-context examples, optional feature selection, and tabular or serialized prompts.

  • I. INTRODUCTION: Biomedical tables are often small, high-dimensional, unordered, and incomplete, making cross-feature interactions difficult to model reliably.These properties distinguish tabular clinical prediction from language and image modeling.
  • I. INTRODUCTION: TAP-GPT adapts TableGPT2 for few-shot classification of Alzheimer’s disease versus cognitively normal individuals from AT[N] feature tables.The framework combines tabular understanding with domain adaptation for multimodal biomarker and imaging-derived data.
  • I. INTRODUCTION: The study evaluates TAP-GPT across QT-PAD, amyloid PET, tau PET, and structural MRI datasets to test generalization beyond one biomarker dataset.The evaluation includes three additional neuroimaging modalities stored in tabular form.
  • I. INTRODUCTION: Feature selection addresses high-dimensional tables, while missingness experiments assess robustness under simulated and naturally occurring incomplete data.The feature-selection strategies include LASSO and LLM-guided ranking under context-length constraints.
  • I. INTRODUCTION: Few-shot prompts combine k labeled, split-specific examples with an unlabeled target patient, without overlap between context pools and targets.The examples are used solely for prompt construction within their corresponding data split.
  • I. INTRODUCTION: TAP-GPT preserves explicit table structure through a semantic table encoder, whereas serialization can lose column alignment and consume more tokens.Its decoder is fine-tuned for diagnosis while the pretrained table encoder remains structurally aware.

A. Datasets

The study uses ADNI data, a public-private resource combining imaging, biological markers, and clinical assessments to measure disease progression.

  • A. Datasets: ADNI combines serial MRI, PET, biological markers, and clinical and neuropsychological assessments to measure progression from mild cognitive impairment and related states.The dataset resource supports multimodal Alzheimer’s disease research.

1) Quantitative templates for the progression of AD Biomarker Data:

QT-PAD provides a multimodal biomarker dataset from ADNI cohorts, covering clinical, PET, CSF, MRI, genetic, and demographic measurements.

  • 1) Quantitative templates for the progression of AD Biomarker Data:: QT-PAD contains multimodal Alzheimer’s progression biomarkers spanning FDG and amyloid PET, CSF measures, structural MRI, APOE4, and demographic covariates.Five cognitive score metrics were excluded because they were highly collinear with diagnosis.
  • 1) Quantitative templates for the progression of AD Biomarker Data:: 333 subjects remained after restricting to baseline measurements without missing values: 237 cognitively normal and 96 Alzheimer’s disease participants.The resulting cohort was extracted from ADNI 1/Go/2 cohorts.
  • 1) Quantitative templates for the progression of AD Biomarker Data:: A second ADNI subset contains participants with available amyloid PET, tau PET, and structural MRI regional summary data.Amyloid-β and tau PET images were processed using standardized pipelines.

2) Regional Summary Imaging Data:

The regional imaging datasets represent amyloid PET, tau PET, and structural MRI through FreeSurfer-derived regional summaries with demographic and genetic covariates.

  • 2) Regional Summary Imaging Data:: FreeSurfer-derived imaging features include 68 cortical and four subcortical regions, yielding 74 regional summaries across the three modalities.The subcortical regions are the thalamus, hippocampus, amygdala, and caudate.
  • 2) Regional Summary Imaging Data:: The imaging tables also include age, gender, years of education, and APOE4 copy number as subject-level covariates.These variables supplement the region-level PET and MRI measurements.
  • 2) Regional Summary Imaging Data:: The experiments compare four prompt formats that cross zero-shot versus few-shot context with tabular versus serialized natural-language representations.Each format predicts Alzheimer’s disease status for the same held-out patient in the illustrative figure.
  • 2) Regional Summary Imaging Data:: Participants were selected using the most recent available visit and restricted to cognitively normal or Alzheimer’s disease diagnoses.This strategy aimed to maximize Alzheimer’s cases and improve class balance.
  • 2) Regional Summary Imaging Data:: Approximately 40% of samples were allocated to training, 10% to validation, and 20% to testing, with remaining samples divided into non-overlapping pools.The splitting strategy keeps prompt-construction data separate from model training, validation, and testing.

3) Creating Tabular Prompts:

The study constructs four patient-level prompt formats that vary contextual exposure and representation, then adapts TableGPT2 into TAP-GPT for binary AD prediction with optional structured reasoning outputs.

  • 3) Creating Tabular Prompts:: In-context pools populate prompts exclusively and are kept separate from model optimization and evaluation.The pools correspond to train, validation, and test splits of approximately 10% each.
  • 3) Creating Tabular Prompts:: Four prompt formats combine zero-shot or few-shot context with tabular or serialized patient representations.Few-shot prompts include labeled examples and an unlabeled target whose diagnosis is masked.
  • 3) Creating Tabular Prompts:: Serialized prompts convert structured features into standardized descriptions, using ROI=value pairs for imaging measurements to preserve organization.Patient covariates precede regional imaging features, and measurement units are retained when applicable.
  • 3) Creating Tabular Prompts:: Interpretability prompts request a JSON prediction, reasoning explanation, and confidence score while keeping patient data and examples unchanged.The supervised objective remains binary classification, so the output-format change does not alter the training loss.
  • B. TAP-GPT Framework: TAP-GPT adapts TableGPT2-7B by combining a static tabular encoder with a Qwen2.5-7B decoder for structured AD prediction.Labeled biomedical tables are used to fine-tune the decoder to map patient features to diagnostic outcomes.
  • B. TAP-GPT Framework: Each patient is converted into all four prompt formats, and a separate model is fine-tuned for each format to isolate structure and context effects.Inputs are generated for training, testing, and validation splits using corresponding in-context pools.
  • B. TAP-GPT Framework: TAP-GPT predicts AD=1 versus CN=0 using supervised learning, fixed-position diagnosis tokens, ten random seeds, and QLoRA parameter-efficient fine-tuning.The fixed token position supports consistent parsing and evaluation across seeds and formats.
  • B. TAP-GPT Framework: Inference constrains predictions to {0,1}, while interpretability experiments add reasoning and confidence outputs without changing the underlying architecture.A custom logits processor enforces valid binary predictions.

C. Experimental Setup

The experiments evaluate TAP-GPT across four prompt formats and multiple baseline categories, using F1-centered tuning and ablations for in-context examples and imaging feature counts.

  • C. Experimental Setup: TAP-GPT is evaluated across four multimodal biomedical datasets and four prompt formats, with F1, balanced accuracy, precision, and recall reported.The prompt formats vary by zero-shot versus few-shot context and tabular versus serialized representation.
  • C. Experimental Setup: QT-PAD k-ablation tests k ∈ {2, 4, 6, 8, 10, 12, 16, 20} across three seeds and selects the most stable high-validation-F1 setting.The selected k is used for subsequent few-shot tabular experiments.
  • C. Experimental Setup: Figure 3 compares QT-PAD mean F1 in zero-shot and k = 8 few-shot contexts, separating tabular and serialized LLM prompts from direct structured-data models.Error bars represent standard deviation.
  • C. Experimental Setup: Imaging feature selection uses training-set-only LASSO before prompt construction, followed by a joint p × k ablation to manage prompt-length degradation.The grid tests p ∈ {8, 16, 32} and k ∈ {4, 8, 12}.
  • C. Experimental Setup: Baselines include classical machine learning, TabPFN, unfinetuned TableGPT2, TableGPT-R1, general-purpose Qwen models, and GPT-4.1-mini.The comparisons assess task-specific fine-tuning, tabular specialization, and stronger general-purpose backbones.
  • C. Experimental Setup: Optuna tunes neural, LoRA, traditional-machine-learning, and imaging feature-count hyperparameters, selecting configurations with consistently strong validation F1 across seeds.The final configurations are evaluated on the test set.
  • C. Experimental Setup: Experiments run as several hundred single-GPU jobs on NVIDIA A100 hardware because modalities, seeds, prompt formats, and ablation combinations create many configurations.Each job uses one GPU, four CPU cores, and one node.

III. RESULTS

TAP-GPT performs strongest in few-shot tabular settings, improves over its backbone across imaging modalities, and generally surpasses traditional baselines while remaining competitive with GPT-4.1-mini.

  • III. RESULTS: The evaluation compares zero-shot and few-shot tabular and serialized prompts across TAP-GPT, classical baselines, TabPFN, and vanilla LLMs.Imaging comparisons cover amyloid PET, tau PET, and structural MRI.
  • III. RESULTS: Zero-shot serialized prompts consistently outperform zero-shot tabular prompts on QT-PAD, while interpretable chain-of-thought prompting improves both formats.The single-row zero-shot tabular format lacks in-context examples.
  • III. RESULTS: Few-shot tabular TAP-GPT and Interpretable TAP-GPT are strongest on QT-PAD, with Interpretable TAP-GPT reaching mean F1=0.89.TAP-GPT outperforms traditional machine-learning baselines, TabPFN, and vanilla LLMs in this task.
  • III. RESULTS: In four-shot imaging experiments with 16 selected features, TAP-GPT consistently outperforms traditional methods and TabPFN.Classical models are disadvantaged by the very low-shot setting of four training samples and one test sample.
  • III. RESULTS: In tau PET, mean F1 rises from 0.43 for Qwen2.5 to 0.48 for TableGPT2 and 0.68 for TAP-GPT.A similar stepwise progression appears in amyloid PET and, less strongly, structural MRI.
  • III. RESULTS: Imaging tabular and structured key-value serialized prompts perform similarly, because key-value formatting preserves organization without substantially increasing prompt length.This contrasts with the QT-PAD serialized few-shot degradation.
  • III. RESULTS: For amyloid PET, Interpretable TAP-GPT achieves F1 0.64 versus GPT-4.1-mini F1=0.65, while tau TAP-GPT reaches F1 0.68 versus 0.53.In MRI, GPT-4.1-mini reaches F1=0.66 and serialized Interpretable TAP-GPT reaches F1=0.65.
  • III. RESULTS: Few-shot prompting consistently improves TAP-GPT over zero-shot in imaging, whereas general-purpose LLMs show smaller gains than on QT-PAD.The authors attribute the improvement to additional task-specific signal from labeled in-context examples.

B. Ablations

Ablations show that optimal in-context example counts vary by model and that imaging performance favors fewer examples and moderate feature counts to control prompt length.

  • B. Ablations: TAP-GPT peaks at k = 8 on QT-PAD, while TableGPT2 peaks near k = 6 and TabPFN improves steadily as k increases.The ablation compares TableGPT2, TabPFN, and TAP-GPT across multiple in-context example counts.
  • B. Ablations: Imaging ablations evaluate mean F1 across combinations of in-context examples k and selected features p over three random seeds.The study tests p ∈ (8, 16, 32) and k ∈ (4, 8, 12) for each modality.
  • B. Ablations: Amyloid PET and tau PET peak at lower k and p, with k = 4 most stable across p and p = 16 pairing best with moderate feature counts.These patterns indicate degradation when large tables are supplied as text input.
  • B. Ablations: MRI shows weaker and more variable gains at larger feature counts or example numbers than amyloid PET and tau PET.The final imaging configuration uses p = 16 and k = 4 across modalities for consistency.

2) k- and p- Ablation Analysis for Imaging Data:

Feature selection improves TAP-GPT’s few-shot imaging performance, while its prompts produce structured, modality-aware explanations that can expose both clinically aligned reasoning and important inconsistencies.

  • Feature dimensionality: Across amyloid PET, tau PET, and MRI, TAP-GPT consistently improves when reducing inputs from 72 regional features to the top 16 selected by LASSO.Using all 72 features causes degradation, particularly for tau PET and MRI.
  • Feature dimensionality: TableGPT-R1 shows near-zero performance with 72 features across modalities, indicating difficulty handling long tabular contexts without feature reduction.Qwen3 largely maintains comparable performance between 72 and 16 features, but remains below TAP-GPT overall.
  • Feature dimensionality: Feature selection enables effective performance for tabular LLMs when clinical tables are high-dimensional.The results suggest that improving the base architecture alone may not resolve long-context challenges.
  • Interpretable prompting: Across four datasets, TAP-GPT returns the diagnosis, probability estimate, and stepwise reasoning in structured JSON outputs.The outputs provide direct insight into how the model synthesizes multimodal clinical features.
  • Interpretable prompting: TAP-GPT integrates genetic risk, amyloid and tau biomarkers, neurodegeneration, and modality-specific imaging evidence in its explanations.Imaging examples include biologically relevant interpretations of hippocampal, amygdala, APOE4, and regional uptake features.
  • Interpretable prompting: Prompt-induced interpretability supports alternative structured outputs without retraining, but explanations can misstate feature directionality or patient-specific values.The authors caution that reasoning strings require validation and should not be treated as causal evidence.

2) LLM-Derived Feature Ranking Across Imaging Modalities:

LLM-derived rankings identify biologically plausible imaging regions while differing partly from LASSO, and TAP-GPT remains comparatively stable under self-reflection.

  • LLM-derived feature ranking: GPT-4.1-mini prioritizes medial temporal, subcortical, and posterior cortical regions in MRI, including hippocampus, amygdala, precuneus, and posterior cingulate.Several prioritized regions are established in neurodegeneration and overlap with LASSO-selected features.
  • LLM-derived feature ranking: Across modalities, GPT-4.1-mini selects a partially distinct subset of regions from LASSO.The authors interpret this as evidence that LLM rankings may provide an alternative or complementary feature-selection strategy.
  • LLM-derived feature ranking: GPT-4.1-mini can serve as a helper agent that ranks features before prompt construction, operating in a similar knowledge space to tabular predictor agents.This use complements the downstream prediction workflow rather than replacing it.
  • LLM-derived feature ranking: Figure 9 maps GPT-4.1-mini’s normalized regional-importance rankings onto anatomical space and labels the top 10 ROIs for each modality.Cortical and subcortical regions are colored by normalized importance scores from 0 to 1.
  • Self-reflection stability: TAP-GPT maintains similar performance between standard prompting and self-reflection across the three imaging modalities.GPT-4.1-mini instead shows considerable performance loss under self-reflection across modalities.
  • Self-reflection stability: Self-reflection stability matters for multi-agent systems because repeated critique, discussion, and voting can otherwise destabilize predictions.The reported single-round result indicates reduced susceptibility to drift during debate.

D. Missingness

TAP-GPT remains robust to both simulated and naturally occurring missingness, with gradual or limited degradation as target-level missingness increases.

  • Simulated missingness: The simulated experiment masks 10%, 20%, 30%, 40%, or 50% of feature values uniformly across each model’s input table.The masking follows a missing-completely-at-random assumption.
  • Real-world missingness: The real-world analysis evaluates 541 patients with naturally occurring missing data in an independent cohort excluded from training and prior evaluation.Twenty percent of this subset was reserved exclusively as an in-context learning pool.
  • Real-world missingness: Real-world results stratify patients by target-level missingness while preserving the natural missingness distribution in the in-context learning pool.The pool’s mean missingness was 0.247, and target-level missingness isolates missing data in the evaluated patient.
  • Real-world missingness: TAP-GPT shows stable performance across real-world missingness bins with limited degradation and relatively low variability across random seeds.TabPFN maintains similar performance to TAP-GPT as missingness increases.

IV. DISCUSSION

Across multimodal and unimodal low-shot experiments, TAP-GPT improves on its backbone, outperforms classical baselines, and remains competitive with other tabular and general-purpose models while providing interpretable outputs and robustness to missingness.

  • Contribution: TAP-GPT introduces a tabular-LLM framework for few-shot AD prediction from multimodal biomedical tables and systematically applies tabular LLMs to biomarker and imaging-derived modalities.The framework addresses low-sample, partially observed, multimodal clinical data without relying on conventional deep learning alone.
  • Performance: Across QT-PAD and three imaging datasets, TAP-GPT shows strong low-shot performance, outperforming classical baselines and improving on its backbone.It remains competitive with dedicated tabular foundation models and general-purpose LLMs, although strong general-purpose models match it in some imaging settings.
  • Performance: Multimodal QT-PAD results are particularly strong, while unimodal imaging results remain competitive.The authors suggest that multimodal analysis could further improve imaging performance.
  • Backbone dependence: Backbone choice materially affects outcomes: specialization improves performance from Qwen2.5 to TableGPT2 and from TableGPT2 to TAP-GPT, while GPT-4.1-mini outperforms TAP-GPT in some imaging experiments.TableGPT-R1 and Qwen3 are not strong in the reported few-shot imaging setting, and TableGPT2’s fixed architecture limits decoder replacement.
  • Low-shot comparison: TAP-GPT is competitive with TabPFN in low-shot settings and often outperforms it in four-shot imaging experiments depending on prompt format.Unlike TabPFN, TAP-GPT combines a tabular encoder with an LLM decoder that can use feature names and semantic knowledge.
  • Robustness: TAP-GPT represents missingness directly in prompts and maintains stable performance under simulated and real-world missingness without explicit imputation.This supports operation on incomplete clinical tables with limited preprocessing.
  • Multi-agent systems: Interpretable multimodal outputs, missingness robustness, and self-reflection stability support TAP-GPT as a candidate for clinical multi-agent workflows.Flexible natural-language or structured explanations can support clinician-facing interpretation and machine-readable communication between agents.
  • Conclusion: The conclusion presents TAP-GPT as groundwork for modality-specific expert models that collaborate toward more reliable and transparent clinical decisions.The reported scope covers few-shot AD prediction on multimodal biomarker and imaging-derived data.
Loading 2603.17191v1…