No Data? No Problem: Robust Vision-Tabular Learning with Missing Values

TL;DR

RoVTL introduces a robust vision-tabular learning framework that remains effective across the full spectrum of tabular data availability. It combines contrastive multimodal pretraining with missingness augmentation, a gated cross-attention fusion, TabMoFe loss, and disentangled gradient learning to maintain strong performance when tabular data are partial or absent. The approach yields state-of-the-art results on UK Biobank CAD classification and BMI regression, generalizes to an external cardiac dataset, and transfers to a 2D natural-vision domain, demonstrating broad applicability. The work advances vision-tabular foundation-model development by addressing edge-case missingness and providing a transferable training pipeline.

Abstract

Large-scale medical biobanks provide imaging data complemented by extensive tabular information, such as demographics or clinical measurements. However, this abundance of tabular attributes does not reflect real-world datasets, where only a subset of attributes may be available. This discrepancy calls for methods that can leverage all the tabular data during training while remaining robust to missing values at inference. To address this challenge, we propose RoVTL (Robust Vision-Tabular Learning), a framework designed to handle any level of tabular data availability, from 0% to 100%. RoVTL comprises two key stages: contrastive pretraining, where we introduce tabular attribute missingness as data augmentation to promote robustness, and downstream task tuning using a gated cross-attention module for multimodal fusion. During fine-tuning, we employ a novel Tabular More vs. Fewer loss that ranks performance based on the amount of available tabular data. Combined with disentangled gradient learning, this enables consistent performance across all tabular data completeness scenarios. We evaluate RoVTL on cardiac MRI scans from the UK Biobank, demonstrating superior robustness to missing tabular data compared to prior methods. Furthermore, RoVTL successfully generalizes to an external cardiac MRI dataset for multimodal disease classification, and extends to the natural images domain, achieving robust performance on a car advertisements dataset. The code is available at https://github.com/marteczkah/RoVTL.

Figures (8)

• Figure 1: RoVTL consistently outperforms image-only inference hasny2025tgv and prior multimodal baselines du2024tip, addressing their key limitation of degraded performance in realistic clinical scenarios where less than 30% of tabular entries are available. The dotted line indicates an edge case evaluated in our work but absent from the original paper du2024tip.
• Figure 2: Overview of RoVTL. Our proposed framework, is composed of two main components: contrastive pretraining and downstream task tuning. In the pretraining stage, we introduce a novel tabular data missingness augmentation strategy, where random tabular entries are dropped (red cross) to teach the encoder to handle missing entries. During downstream task tuning, two subsets of the tabular attributes $t$ are used as input into the encoder and ranked by the proposed TabMoFe loss. Further, our multimodal fusion module weights the importance of the tabular data for the mulimodal fusion. Not illustrated in the figure is DGL wei2025boosting, which is integrated into our fine-tuning pipeline.
• Figure 3: Model performance under different data availability scenarios from 0% to 100% tabular data used as input. CAD is evaluated using AUC ($\uparrow$), BMI using MAE ($\downarrow$), and DVM model classification using accuracy ($\uparrow$). : frozen backbones, : trainable backbones.
• Figure 4: The effect of different tabular augmentation techniques used at pretraining on the result of CAD classification under random value missingness. : frozen, : trainable backbones.
• Figure 5: Model performance plotted over tabular data availability from 0% to 100%. CAD is reported using AUC ($\uparrow$), BMI using MAE ($\downarrow$), and DVM car classification using accuracy ($\uparrow$). For clarity, zoomed-in versions are provided for BMI regression and car model classification with DVM to higlight the performance differences between the models. : frozen backbones, : trainable backbones.