List of Papers Browse by Subject Areas Author List
Abstract
Medical tabular data are ubiquitous in clinical research, but deep learning for tables remains underexplored because reliable labels often require costly expert adjudication, even though structured clinical variables are routinely available in tabular form. Self-supervised learning can leverage these unlabeled tables, and recent binning-based pretexts offer a promising inductive bias, but existing objectives fix a single global quantile discretization and apply feature-agnostic supervision. We propose Adaptive Binning, a training-adaptive discretization pretext for tabular SSL that couples discretization to learning through a feature-wise coarse-to-fine curriculum. Motivated by the spectral bias of neural networks and the principles of curriculum learning, our method progressively refines discretization per feature upon plateau detection and selects representation-aware splits to jointly improve value-space concentration and representation-space coherence. A heterogeneity-aware objective unifies categorical reconstruction with ordinal supervision for numerical features, and experiments on public medical tabular datasets under unified evaluation protocols show consistent gains for linear probing and fine-tuning without dataset-specific discretization tuning. We further introduce a medical tabular SSL benchmark with standardized protocols to support reproducible progress in this underexplored domain. Our code is available at https://github.com/labhai/Adaptive-Binning.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4808_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/labhai/Adaptive-Binning
Link to the Dataset(s)
N/A
BibTex
@InProceedings{KimDae_When_MICCAI2026,
author = { Kim, Daehwan AND Chung, Haejun AND Jang, Ikbeom},
title = { { When, Where, and How: Adaptive Binning for Tabular Self-supervised Learning } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
Proposes an Adaptive Binning framework that replaces the conventional fixed global binning used in SSL for medical tabular data with a per-feature, coarse-to-fine discretization that evolves adaptively during training. The framework comprises three modules (FPT, DIGS, HORD) and is accompanied by a unified benchmark over eight public medical tabular datasets.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
-
Intuitive and clean framing The problem is structured around three questions — When / Where / How — so the role and motivation of each module are clearly articulated.
-
Novelty A tabular SSL method that adaptively adjusts discretization during pretraining, with a representation-space-based criterion for bin splitting.
-
Generality across encoders Table 4 evaluates fine-tuning on five encoders (MLP, ResNet, TabNet, FT-Transformer, T2G-Former), confirming an encoder-agnostic effect.
-
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
-
Poor clarity and readability (major concern) The formulation and overall presentation are difficult to follow. The paper should present the method more clearly, with concise explanations of each component and its role.
-
Insufficient empirical evidence for the core adaptive binning claim Although adaptive binning is the central contribution, the paper does not provide direct evidence of how the binning evolves (e. g. , no bin-schedule visualization or bin statistics).
-
Ambiguity in ablation design FPT and DIGS appear mutually dependent, yet the ablation study does not clearly describe how components are removed or substituted. This makes it difficult to interpret the contribution of each module.
-
Incomplete and inconsistent evaluation Fine-tuning is reported on only a subset of datasets (4/8) without justification, and model design choices (e. g. , MLP variants) vary across datasets without explanation. This limits the reliability of the evaluation.
-
Missing analysis of computational overhead The additional cost introduced by DIGS is not discussed, leaving the practical efficiency of the method unclear.
-
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Positive factors: Clear and well-motivated When/Where/How framing; novel adaptive, representation-aware discretization for tabular SSL; broad hyperparameter robustness sweep; encoder-agnostic effectiveness across five architectures; unified benchmark over eight public medical tabular datasets. Negative factors: Core adaptive claim is not empirically visualized (no bin-schedule evolution; bin counts unspecified for both method and baselines); ambiguous ablation protocol due to mutual dependence between FPT and DIGS; fine-tuning reported on only 4 of 8 datasets without justification; no computational-overhead analysis for DIGS; per-dataset MLP design choices left unexplained. Overall: The paper offers genuine novelty and a reasonable formulation, but suffers from insufficient empirical support for its core claim and a shallow interpretation of the experimental results. Adding bin-schedule visualizations, clarifying the ablation substitutions, and running consistent fine-tuning across all 8 datasets with corresponding analysis would justify an upgrade in score.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
My concerns have been adequately addressed in the rebuttal. I find the proposed method meaningful and novel.
Review #2
- Please describe the contribution of the paper
This paper proposes Adaptive Binning, a training-adaptive discretization pretext task for tabular self-supervised learning. It is designed to address the limitation of using a single global quantile discretization in prior work. The method progressively refines feature-wise discretization when training plateaus and performs representation-aware split selection. The paper also introduces a heterogeneity-aware objective that combines categorical reconstruction with ordinal supervision.
The method is evaluated on multiple tabular datasets and reports favorable improvements over prior SOTA baselines.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
S1: The paper studies tabular representation learning, which is an important topic for medical AI applications.
S2: The adaptive discretization design is well motivated and technically reasonable.
S3: The reported performance gains are meaningful, and the experimental section is relatively comprehensive.
S4: The manuscript is generally clear, well organized, and easy to follow.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
W1: Some existing SSL approaches model heterogeneous tabular features separately with regression/classification objectives (e.g., MSE for numerical features and cross-entropy for categorical features) [R1, R2], without relying on discretization. The paper should provide a clearer conceptual or empirical comparison against this line of methods, including when adaptive binning is preferable.
W2: There are also methods with learnable discretization/tokenization mechanisms that aim to preserve ordinal information [R3]. The novelty and practical advantage of Adaptive Binning over such approaches are not yet fully clarified.
[R1] Du, Siyi, et al. “TIP: Tabular-image pre-training for multimodal classification with incomplete data.” European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024. [R2] Somepalli, Gowthami, et al. “SAINT: Improved Neural Networks for Tabular Data via Row Attention and Contrastive Pre-Training.” NeurIPS 2022 First Table Representation Workshop.
[R3] Yan, Jiahuan, et al. “Making Pre-trained Language Models Great on Tabular Prediction.” The Twelfth International Conference on Learning Representations.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I find the problem important and the proposed idea interesting, and the empirical results are promising. However, the current draft does not yet provide sufficiently clear positioning against some closely related directions. Clarifying these differences through more direct comparisons and discussion would make the paper better.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I find the rebuttal convincing overall. The authors have largely addressed my main concerns by clarifying the distinction from prior methods, and by adding empirical comparisons with SAINT and tree-based baselines.
I also find the responses to the other reviewers’ concerns helpful. In particular, the added bin evolution statistics addresses the main concern about whether adaptive binning actually occurs during training. I agree that clinical tabular data is a valid and important medical data modality for MICCAI. While broader transfer experiments and additional discussion would further strengthen the paper, I do not view the absence of imaging integration as a fundamental relevance issue.
Review #3
- Please describe the contribution of the paper
Feature-wise plateau trigger (FPT) for training-adaptive, per-feature discretization refinement in tabular SSL Dispersion-Informed Gain-based Splitting (DIGS) coupling value-space variance reduction with representation-space coherence for split selection Heterogeneity-aware ordinal loss (HORD) unifying categorical cross-entropy with soft ordinal supervision for numerical features Medical tabular SSL benchmark with standardized protocols across 8 datasets and 4 task types
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
Well-structured method with clear, orthogonal components validated by strong ablations Rigorous hyperparameter analysis with statistical significance across datasets Consistent transfer gains across multiple downstream architectures Useful benchmark with standardized protocols improving reproducibility
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
Limited relevance to MICCAI, as the work is purely tabular with no imaging or clinical workflow integration Missing comparisons with strong tabular baselines (both SSL methods and tree-based models like XGBoost/CatBoost) SSL motivation is weak, given the small dataset sizes Evaluation is restricted to in-dataset pretraining without demonstrating transfer learning benefits The practical advantage over standard tabular methods is not clearly established
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
The paper’s strongest contribution for a clinical AI venue is the benchmark (Table 1) with standardized protocols — this alone could support a workshop paper but is insufficient as the primary MICCAI contribution. The spectral bias motivation for coarse-to-fine curriculum is conceptually appealing but never empirically verified — no experiment shows that coarser bins are learned first or that refinement timing correlates with spectral learning dynamics. Results on HFC (299 samples) consistently show no benefit from FPT/DIGS (Table 3 w/o FPT: same as Ours on HFC) — this suggests the method degrades gracefully on small datasets but also raises the question of whether the improvements on larger datasets are driven by dataset scale rather than the adaptive mechanism itself.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Technically sound with clear design and strong empirical analysis, but limited MICCAI relevance and incomplete baseline comparisons weaken the overall impact. The SSL motivation and evaluation setup are not fully convincing, making the contribution insufficient for acceptance.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank all reviewers for their constructive comments and recognition of our novel training-adaptive binning pretext, unified medical tabular SSL benchmark, clear When/Where/How design, and strong, encoder-agnostic gains. R#1
- Clarity: We will add Sec.2.2 pseudocode to reduce text-heavy presentation, with concise explanations of each component and its role.
- Bin evolution: Refinement occurs in 7/8 datasets; median per-dataset mean bin count increases 5.54→17.44.Adaptation is thus observed, and we will add bin-schedule visualizations/statistics, including BR’s fixed T.
- Ablation: FPT & DIGS are not mutually dependent (Tab.3): w/o-FPT and w/o-DIGS each correspond to When & Where. w/o-FPT keeps DIGS with fixed-epoch refinement; w/o-DIGS keeps FPT with value-space variance reduction only. Tab.3 caption will be revised for clarity.
- overhead: DIGS affects pretraining only, not fine-tuning/inference; mean wall-clock overhead across pretraining is 7.48%.
- Fine-tuning coverage: Tab.4 was task-stratified, not performance-selected, with one dataset per task (BC/NMC/OMC/Reg) due to space; we will expand it to all datasets/encoders. MLP fine-tuning on omitted datasets gives Ours 1st on ILPD↑, CTG↑, PT↓, and 2nd on MHR↑: 74.9 vs Sup 74.5; 87.6 vs BR 87.3; 13.0 vs BR 13.3; 63.0 vs BR 63.9.Design choices for MLP variants (Tab.1) are set per dataset via supervised validation and fixed across Tabs.2-4 for reliability; non-MLP encoders use defaults.
R#2&3
- Baselines: Adaptive Binning is training-adaptive SSL, not discretization alone. Unlike [R1-2 by R#2] heterogeneous-feature regression/classification or [R3] learnable discretization/tokenization, it creates label-free ordinal targets, uses HORD, and refines them via FPT/DIGS. As requested, we compare on Tab.4 datasets against strong tabular baselines—tabular SSL (SAINT [R2]) and tree-based models—with Ours’ best encoder per dataset. SAINT-Sup/SAINT-SSL/XGBoost/CatBoost/Ours: HFC↑ 92.3/92.5/92.2/92.4/94.2; ESR↑ 71.8/74.7/70.7/67.1/76.9; EOL↑ 95.5/94.9/95.0/91.7/97.1; BFP↓ 4.40/4.51/4.86/4.85/4.77.Ours is best on 3/4 vs SAINT, 4/4 vs trees. Its practical role is SSL initialization for numerical and graded clinical variables, not replacing supervised tabular learners.
R#3
- MICCAI relevance: Purely tabular focus does not limit MICCAI relevance. We do not claim current imaging/clinical workflow integration; we target label-scarce, workflow-native clinical tables, not generic tabular data. Because medical SSL remains imaging/language-centric, label-efficient SSL for this modality is a clinical need, not a convenience. Our contribution is twofold, coupling Adaptive Binning, the primary methodological advance for clinical tabular SSL, with a new benchmark for reproducible evaluation. This scope is consistent with MICCAI’s call for papers on integration of imaging/non-imaging and multimodal data, instantiated at MICCAI’25 by ≥6 non-imaging papers, e.g., ProCDS (Tan et al.) and MVP-LLMs (Hao et al.).
- Motivation/transfer: For small-dataset SSL motivation and in-dataset-only evaluation, we provide semantically matched EOL→BFP transfer with reinitialized input/head: unlabeled EOL pretraining improves BFP↓ at all fractions. Sup/Transfer at 25, 50, 75, 100%: 5.69/5.39, 5.92/5.47, 5.61/5.19, 5.50/5.07—transfer beyond in-dataset pretraining; broader transfer is Sec.4 future work.
- Spectral bias: Two EOL analyses test whether coarser bins are learned first and refinement timing follows the coarse-to-fine curriculum. (i) To reach 95% final performance, coarse targets (2/4 bins) needed ≤10 epochs and fine targets (8/16 bins) 400/500 epochs. (ii) Refinement was first accepted at epoch 18, consistent with this ordering.
- HFC/scale: HFC is a no-refinement case, not evidence of scale-driven gains. Because refinement never triggers, Ours matches w/o FWA/FPT/DIGS in Tab.3.The smaller BFP (n=252) did refine, and w/o FWA/FPT/DIGS worsened RMSE; gains reflect the adaptive mechanism, not scale alone.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
A further rebuttal is required for the reviewers to reach a consensus.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Great work.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
While the reviews are still mixed, the overall consensus ultimately favors acceptance. The authors provided a comprehensive rebuttal that effectively addressed the core technical critiques by expanding baseline comparisons and clarifying the underlying mechanics of the method. Ultimately, the robust empirical validation and the practical value of the proposed framework represent a good contribution to the community.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The presented work is novel and provides a valuable benchmark for future research in medical tabular self-supervised learning. Its adaptive discretization design is well motivated and technically reasonable, with several strengths broad hyperparameter robustness analysis, encoder-agnostic effectiveness across five architectures, and a unified benchmark over eight public medical tabular datasets. The concerns raised during review were well addressed by the authors in the rebuttal. One remaining concern relates to MICCAI relevance, since the paper focuses on non-imaging medical tabular data. While the authors argue that MICCAI does consider non-imaging data, my understanding is that such cases are typically strongest in multimodal settings where at least one modality is medical imaging. Nevertheless, aside from this scope concern, the paper is technically solid and valuable. Therefore, I recommend acceptance.
