Abstract
Recent advances in deep learning for tabular data (TabM, SAINT, TabPFN) have demonstrated strong performance on curated benchmarks, but their behaviour on real-world clinical datasets remains poorly understood. Clinical records present challenges that benchmarks do not, such as high feature missingness, small positive-class counts, multi-center heterogeneity, and domain-specific correlations that differ from standard benchmark distributions. We benchmark eight tabular encoders, four differentiable (multi-layer perceptron, TabM, SAINT, TabPFN) and four classical (logistic regression, random forest (RF), gradient boosting (GB), XGBoost), within a shared multi-modal framework that fuses preoperative computed tomography angiography (CTA) of vascular anatomy with clinical tabular data for reintervention risk prediction after Endovascular aneurysm repair (EVAR). The framework uses Merlin, a 3D radiology foundation model, for visual encoding, a learned L1-sparse feature gate for clinical inputs under high missingness, and featurewise linear modulation-based cross modal fusion. Experiments on the RADAR consortium database show that RF and GB near-perfectly memorise training data (AUROC 1.000) but suffer large test-set drops (gaps of 0.343 and 0.331), while differentiable encoders show smaller
train/test gaps (below 0.09 for TabM, SAINT, and TabPFN). XGBoost achieves the best test AUROC overall (0.711), while GB achieves the best F1 and specificity among classical methods. These results expose a tension between discriminative capacity and generalization that is absent from standard tabular benchmarks, with direct implications for clinical deployment.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MultiTab_017.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=4bCFVt14My
BibTex
@InProceedings{RanAmi_Tabular_MICCAISAT2026,
author = { Ranem, Amin AND Jebbink, Erik G. AND Yeung, Kak Khee AND Geelkerken, Bob H. AND Zeebregts, Clark J. AND Reijnen, Michel M. P. J. AND Wolterink, Jelmer M.},
title = { { Tabular Encoder Generalization Under Clinical Constraints for Multi-Modal Risk Prediction } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17263},
month = {pending},
page = {pending}
}