List of Papers Browse by Subject Areas Author List
Abstract
Age-dependent performance disparities in medical image classification often arise because age acts as a confounder, linking imaging morphology with disease prevalence. In practice, disparities can manifest as overdiagnosis at ages where disease prevalence is higher and underdiagnosis at ages where prevalence is lower, and can become more pronounced under train–test shifts in the age distribution. Conventional mitigation approaches that enforce strict age invariance may suppress diagnostically meaningful information encoded in age. We therefore propose a robust framework that mitigates the effects of age-dependent confounding by targeting first-order, label-conditioned age–error trends rather than enforcing invariance. Following a warm-up phase, we characterize model-perceived difficulty and model its age-dependent trends in a label-conditioned manner. We decorrelate age from dominant age–difficulty trends using robust, Huber-weighted affinity weights, attenuating confounding driven shortcuts while preserving clinically meaningful, non-linear age information. We further introduce an Age Coverage Score that scales the decorrelation penalty by mini-batch age variance to ensure stable optimization under limited age diversity. Across two radiology datasets, our approach reduces age-dependent true- and false-positive disparities with minimal AUC impact and remains robust to increasing train–test age distribution shifts.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2145_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{KurNik_Robust_MICCAI2026,
author = { Kurian, Nikhil Cherian AND Parra, Victor Caquilpan AND Shoby, Abin AND Whitbread, Luke AND Palmer, Lyle J.},
title = { { Robust Mitigation of Age-Dependent Confounding Effects via Sample-Difficulty Decorrelation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16896},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors propose a novel bias mitigation method for the specific problem of age-related bias in disease diagnosis models. The method fundamentally works by (differentiably) penalizing FPR/TPR trends across the age spectrum, with specific mechanisms implemented to address batch variability in age distribution. In two CXR diagnosis experiments, the new method outperforms various baseline methods both on bias mitigation alone as well as in terms of the AUC-fairness trade-off.
- 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.
The age-related bias problem is widespread, important, and not yet extensively analyzed or solved. A strong solution to this problem would be very attractive.
The proposed mitigation method is well-motivated, thoroughly developed, and - to my knowledge - rather innovative and quite distinct from any other bias mitigation approaches that I know of.
One particularity concerns the fact that the bias attribute (age) is really treated in a continuous fashion here, whereas most other approaches inherently rely on categorical bias variables. There is very little work on bias mitigation w. r. t. continuous attributes yet.
The baseline methods are well-chosen and quite comprehensive; the experiments are well-designed, metrics carefully and sensibly chosen, and the new method indeed outperforms all of the competitor methods quite substantially on all relevant dimensions.
The manuscript is well-written and the figures are well-designed.
- 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.
While there is not much prior work on bias mitigation for continuous attributes, there is some, and this is currently not referenced or discussed. A few examples that I know of:
- https://proceedings. mlr. press/v97/mary19a/mary19a. pdf
- https://link. springer. com/chapter/10.1007/978-3-030-86520-7_46
- https://openreview. net/pdf?id=YigKlMJwjye
- https://dl. acm. org/doi/abs/10.1109/TPAMI. 2025.3538915 Including one or several of these methods as additional baselines would further strengthen the manuscript, since these are really the most appropriate baselines. (I will be honest - I have not worked with these myself and do not know how tricky they are to get to work. )
- 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
I have just a few remarks that the authors may or may not wish to incorporate into their already very good manuscript.
At the bottom of page two where the identity gi=|pi-yi| is introduced, I had to guess a bit from context what some things meant. An explicit “gi = | d BCE / dlogit |” would have helped me. (Adding a specific page number to the reference to the Murphy book might also help. )
From the main text, it is not immediately obvious that “MIMIC” here means “MIMIC-CXR”.
Some of the texts e. g. in Fig. 3 are too small.
The authors make the case that making the latent space explicitly invariant w. r. t. the sensitive variable of interest may be undesirable. They may find this piece interesting that explicitly discusses this very issue: https://arxiv. org/abs/2305.01397
The authors go to rather great lengths to mitigate minibatching-related issues. Would it not be an alternative approach to design a dataset-level version of the same loss that buffers results from other batches and thus achieves more stable estimates of the age-bias slope? Notice this can still be implemented in a fully differentiable manner. (Obviously fine to leave this for future work, just a suggestion. )
It would be interesting to discuss potential applications of the new mitigation method to other settings beyond age bias.
Finally, for future work, it would be interesting to analyze the effects on generalizability in real-world datasets, e. g. via o. o. d. performance, impact on performance disparities across other sensitive variables, etc.
- 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?
Relevant problem (age bias), interesting and truly novel bias mitigation approach, addresses rather under-researched area (bias mitigation for continuous attributes), good experiments/metrics/baselines, very promising results, well written & motivated. I can find very little to criticize here.
- Reviewer confidence
Very confident (4)
- [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 have read the other reviews and the authors’ rebuttal, and my assessment remains unchanged. This is a really thorough and well-done methodological contribution to the bias mitigation field. I can see nothing in this paper or the rebuttal process that would warrant a rejection.
Review #2
- Please describe the contribution of the paper
The paper introduces a regularization term that enforces equal loss across age bins, achieving better performance on age-separation metrics and comparable AUC to other methods on datasets with linear age-ce-loss relations.
- 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.
- The paper has a well-established experiments section that (while limited to two datasets) explores different distributions and compares to different existing methods.
- The motivation for the contribution is well motivated.
- 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.
Minor Points and Questions
1.Abstract claim on distribution shift. The abstract states that the observed effects “may worsen” under train-test distribution shifts. We would suggest more precise language: the effects do not worsen per se, but rather become more dominant across the data. We ask the authors to revise this characterization accordingly.
2.Motivation for using the gradient as a dynamics metric. In the binary cross-entropy setting, the gradient correlates directly with prediction error, which is itself the quantity of practical interest. It is unclear what additional value the gradient-based framing provides over simply using prediction error. We ask the authors to clarify the motivation for this choice.
3.Warm-up period. The paper does not sufficiently motivate the introduction of a warm-up period. We ask the authors to provide a principled justification or ablation demonstrating its necessity.
4.Age Coverage Modulation and label sensitivity. Since beta is computed with respect to the label, it would be natural for the Age Coverage Modulation to also be defined with respect to the labels present in the minibatch. As currently formulated, the score lacks the label sensitivity needed to be fully expressive. We ask the authors to address this discrepancy.
5.Notational clarity. The paper would benefit from explicitly stating beta = Cov/Var as a standalone equation, which would significantly aid reader intuition. Additionally, the inconsistent switching between subscripts y_i and y is a source of confusion and should be harmonized throughout.
6.Missing evaluation of slope steepness. The experimental section lacks an analysis of the steepness of the calibration slopes. Both MIMIC and CheXpert exhibit approximately similar slopes even under distribution shift, and this should be evaluated and discussed.
7.Terminology: prediction error vs. sample difficulty. The paper at times conflates current prediction error with sample difficulty. These are distinct concepts in the established literature. We strongly encourage the authors to adopt more precise language: what is being measured is the model’s current prediction error on a given sample, not an intrinsic notion of sample difficulty.
Major Points
1.Circular dependence in the estimation of beta. The estimates of beta rely on weights w_i, which in turn depend on the residuals, which themselves depend on beta. This circular dependence makes it difficult to understand how the regularizer is computed in practice. We ask the authors to provide a clear, step-by-step description of the computation, ideally specifying at which point in training each quantity is evaluated and whether any fixed-point iteration or approximation is involved.
2.Relationship to Zafar et al. (2017). In essence, the proposed regularizer enforces equalized losses across different age groups. While the paper does not make this connection explicit, the method can be understood as a direct, normalized extension of the fairness constraints introduced by Zafar et al. (2017, arXiv:1507.05259). We consider this a significant omission and ask the authors to discuss this relationship and clearly delineate their methodological contribution relative to this prior work.
3.Limitation to the linear setting. The paper does not adequately address the fact that the proposed method is designed for and limited to the linear case. This is a substantially weaker setting than that addressed by methods capable of correcting non-linear trends, such as Baharlouei et al. (2024, arXiv:2312.03259). We ask the authors to clarify in which specific settings their method offers advantages over existing approaches, and to provide an honest discussion of where and why it falls short in the non-linear regime.
4.Insufficient discussion of limitations. The current discussion of the method’s limitations is cursory. A thorough treatment should address, at minimum, the restriction to linear settings, sensitivity to hyperparameters such as the warm-up period, and the assumptions underlying the coverage modulation. We ask the authors to expand this section substantially.
- 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 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.
(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?
The paper does not extend substantially over existing methods and is limited to a specific bias setting.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The method introduces a slope-regularized objective with the slope being determined by the sample difficulty is approximated using sample level gradients. The claim is that the introduces a bias regularization while retaining meaningful training signal. Yet the incorporated objective is itself a proxy of what is meaningful information (i.e. the gradient signal). The true impact can not really be differentiated from spurious contribution without varying the regularization magnitude delta and investigating the effect. Moreover the requirement of such a delta indicactes that the method itself is contradicting the initial learning objective and needs careful tuning.
Review #3
- Please describe the contribution of the paper
This paper proposes a robust framework for mitigating age-dependent confounding effects in medical image classification, addressing the critical issue of models learning spurious age-related trends influenced by the age distribution. The main contribution is the introduction of label-conditioned sample-difficulty decorrelation strategy that targets dominant linear age-difficulty trends without enforcing age invariance, thereby preserving diagnostically relevant age-related information. Additionally, the framework further introduces trend affinity weights derived from Huber regression to robustly downweight abnormal samples during decorrelation, and an Age Coverage Score that adaptively scales the regularization penalty based on mini-batch age diversity to ensure stable optimization under limited age representation. Extensive experiments on two chest X-ray datasets under varying age distribution shifts demonstrate that this method outperforms existing fairness mitigation baselines in both effectiveness and robustness.
- 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.
Firstly, this paper targets the critical challenge of age-dependent confounding in medical image classification, and proposes a robust decorrelation framework that models the label-conditioned linear correlation between age and sample difficulty. This approach effectively addresses a key limitation of existing methods that enforce age invariance. Secondly, the framework incorporates robust and adaptive components, including the Huber regression-based trend affinity weights enhance resilience to clinical variability and annotation noise, and the Age Coverage Score effectively downweights decorrelation updates when age coverage is limited, improving robustness under age imbalance and train-test age distribution shifts. Thirdly, experiments cover multiple age distribution shift intensities, two distinct radiology datasets, and comparisons with state-of-the-art baselines, providing clear evidence of the method’s superiority in balancing fairness and diagnostic accuracy.
- 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.
Firstly, the paper uses a one-epoch warm-up phase to estimate trend affinity weights, and immediately fixes these weights after the warm-up phase without any updates during subsequent training. Is this design reasonable? Because as the model learns, the age-difficulty trend may change, which could make the initial weights outdated. Secondly, the adoption of the median absolute deviation (MAD) as the robust scale estimator δ_{y_i} for trend affinity weight calculation is not substantiated by comparative experiments. The paper fails to evaluate alternative robust estimators to verify whether MAD is optimally suited for handling outlier samples in diverse medical image classification. Thirdly, the regularization coefficient λ in Formula (8) is set to 1.2 without a parameter sensitivity analysis, and the paper provides no guidance on adapting this hyperparameter across different clinical datasets.
- 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 does not provide sufficient information for 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.
(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?
Please see the major weaknesses
- 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 have read the response, and it has addressed most of my concerns. Therefore, I recommend accepting the paper.
Author Feedback
We thank R1,2,3 for the useful reviews & for recognizing the critical need to mitigate age confounding effects under distribution shift.
[R2] Prior works: Zafar et al. impose a covariance constraint between sensitive attribute and signed decision-boundary distance as a proxy for decision-boundary unfairness. Our goal differs: continuous-age EOdds under train-test age shift. Age can be clinically informative; thus we mitigate dominant age-error trends learned as spurious correlations, not all age effects- avoiding loss of useful age information. In our method, the covariance/variance form is not an explicit fairness constraint; it arises naturally when estimating label-specific age-error slopes beta_hat_y for g_i = |p_i - y_i|. Keeping beta_hat_y near zero maintains flat label-specific error trends across age, improving EOdds robustness when train-test age distributions differ. Huber affinity & age coverage stabilize estimation. We will discuss this distinction.
[R2] Gradient/error formulation: We will clarify g_i as the logit-gradient magnitude[R1]. For BCE (sigmoid), g_i = |dL_BCE/dlogit| , so g_i = |p_i - y_i| in [0,1]. Estimated post-warm-up, before memorization or gradient shrinkage dominates [Arpit et al., ICML 2017], g_i is a bounded proxy for model-perceived difficulty. Further, we prefer this over p_i since p_i is not label-conditioned. Raw BCE loss is excluded as it is unbounded.
[R2,3] Design choices - coverage, beta, warm-up: We will clarify our method step-by-step. After warm-up, label-conditioned Huber regression estimates trend affinity weights w_i on the full training set; these weights are then fixed. During training, only g_i evolves, and beta_hat_y is a standard WOLS estimate of g_i against fixed w_i & z_i; no fixed-point iteration is used. Fixing w_i is motivated by the fact that early-epoch gradients best reflect each sample’s alignment with the dominant age-difficulty trend before memorization and gradient shrinkage dominate [Arpit et al., ICML 2017]; before warm-up, g_i reflects random initialization, making trend estimation unreliable. One epoch is a viable intervention point: with ~40k training samples a single epoch spans many iterations - enough to yield a stable signal - while still preceding spurious age-shortcut memorization. We acknowledge that warmup epoch =1 is a hyperparameter and will note a sensitivity study. Updating w_i during training would introduce circular co-dependence between trend affinity weights and beta, risking optimization instability. Coverage modulation is already label-conditioned via C_y in Eq. 8; we will make this more clear.
[R2,3] Linearity, generalization & limitations: We agree the method targets the dominant first-order linear age-error trends and not all age effects based on empirical observations (Fig. 1). We flatten the dominant main age-shortcut mode without enforcing distributional/feature invariance [Baharlouei et al.], since nonlinear or higher-order age effects may encode clinically meaningful morphology/prevalence signals (introduction). The framework is most effective for cohorts with monotonic age-prevalence trends, e.g., adult cohorts where chronic disease prevalence often increases with age. A zero-slope target also supports robustness under train-test age shift, since it is independent of the sampled age distribution. Nonlinear residual trends remain outside current scope; spline/kernel extensions are natural future work when such modes dominate. -MAD was used as a standard robust scale estimator for Huber residual weighting, not as uniquely optimal; IQR/SD alternatives will be noted as future work. We will expand the limitation section to cover these points.
Lambda selection: Lambda controls the fairness-accuracy trade-off and was selected by validation grid search maximizing the validation ratio of DeltaSep10 reduction to AUC drop relative to ERM. These details will be added
All remaining minor discussion points will be addressed.
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.
This paper addresses the important problem of age-dependent confounding in medical image classification. The reviewers generally recognized the paper’s strong motivation, methodological innovation, and coherent experimental evaluation. The paper’s significance in continuous-attribute bias mitigation would be better highlighted if the authors address the reviewers’ concerns, including strengthening the positioning relative to prior work, justifying the gradient-based formulation, clarifying key design choices, providing component motivation and sensitivity analyses, and discussing the method’s limitations and generalizability.
- 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 rebuttal appears to have addressed the reviewers’ concerns directly and helped clarified the distinction between this method and existing continuous-attribute bias mitigation approaches. The combination of significance, technical focus, and a constructive rebuttal justifies acceptance.
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.
The rebuttal addresses most of the concerns. One reviewer still raises a valid concern about the regularization parameter and the need to better analyze its effect. However, the other reviewers recommend acceptance, and the authors clarify several important points, including the computation of beta, the relation to prior work, and the linear scope of the method.
I recommend acceptance. In addition to the other modifications mentioned in the rebuttal, the authors should revise the final version to clearly narrow the method’s scope to dominant first-order linear age-error trends. They should also include or strengthen the analysis of the effect of delta, and better explain how the proposed regularized objective differs from and interacts with the original learning objective.
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.
I would like to raise a point regarding the authors’ claim that “the model over-relies on age as a predictive shortcut.” The evidence presented in the paper does not fully support this statement. What the experiments demonstrate is that age has a measurable impact on prediction error. However, this does not necessarily imply that the model is using age as a shortcut in the causal or algorithmic sense. The distinction is important and should be clarified more carefully in the manuscript.
That said, I still find the proposed approach interesting and potentially valuable, as it addresses an important and timely problem: mitigating age-related bias while preserving age-related information that may be informative for the prediction task. This is a challenging problem that has received increasing attention in the neuroimaging and medical AI communities.
While the authors do not fully demonstrate that their method achieves this objective, the proposed framework is nevertheless promising. In particular, the behavior of the method appears to depend strongly on the choice of the threshold parameter (\gamma). Depending on its value, the approach may either remove too much age-related information or fail to remove enough of the age effect. A more detailed analysis of this trade-off would strengthen the paper considerably.
In addition, a more comprehensive comparison with established age-correction and harmonization approaches, such as W-score normalization and ComBat-based methods, would have been valuable. These methods represent important baselines for the problem under consideration and would help position the contribution more clearly with respect to the existing literature.
Nevertheless, despite these concerns, two of the three reviewers support acceptance, and I concur with their overall assessment. The paper tackles a relevant problem and proposes an interesting direction that may stimulate further work in this area.
I strongly encourage the authors to address, in the final version, the majority of the points raised by Reviewer 2, which are largely requests for clarification rather than fundamental methodological objections. Doing so would significantly improve the clarity and positioning of the manuscript.
