List of Papers Browse by Subject Areas Author List
Abstract
Tool-using medical agents can improve chest X-ray question answering by orchestrating specialized vision and language modules, but this added pipeline complexity also creates new pathways for demographic bias beyond standalone models. We present DUCX (Decomposing Unfairness in Chest X-ray agents), a systematic audit of chest X-ray agents instantiated with MedRAX. To localize where disparities arise, we introduce a stage-wise fairness decomposition that separates end-to-end bias from three agent-specific sources: tool exposure bias (utility gaps conditioned on tool presence), tool transition bias (subgroup differences in tool-routing patterns), and model reasoning bias (subgroup differences in synthesis behaviors). Extensive experiments on tool-used based agentic frameworks across five driver backbones reveal that (i) demographic gaps persist in end-to-end performance, with equalized odds up to 20.79%, and the lowest fairness-utility tradeoff down to 28.65%, and (ii) intermediate behaviors, tool usage, transition patterns, and reasoning traces exhibit distinct subgroup disparities that are not predictable from end-to-end evaluation alone (e.g., conditioned on segmentation-tool availability, the subgroup utility gap reaches as high as 50%). Our findings underscore the need for process-level fairness auditing and debiasing to ensure the equitable deployment of clinical agentic systems. Code is available at https://github.com/Nanboy-Ronan/DUCK.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2160_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Nanboy-Ronan/DUCK
Link to the Dataset(s)
N/A
BibTex
@InProceedings{XuZik_DUCX_MICCAI2026,
author = { Xu, Zikang AND Jin, Ruinan AND Li, Xiaoxiao},
title = { { DUCX: Decomposing Unfairness in Tool-Using Chest X-Ray Agents } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16882},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper’s main contribution is to move fairness evaluation for medical agentic systems beyond final-answer metrics and into the internal decision process of the agent. Concretely, it proposes DUCX, a stage-wise fairness decomposition for MedRAX-style chest X-ray QA agents that separates end-to-end disparity into three components: tool exposure bias, tool transition bias, and LLM reasoning bias.
- 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.
1.Novel decomposition of fairness for tool-using agents 2.The evaluation spans five driver backbones, two demographic axes (gender and age), and two datasets, including CheXAgentBench and the curated MIMIC-FairnessVQA. The paper does not stop at aggregate accuracy; it reports multiple fairness metrics such as ΔACC, DP, EoD, and FUT, and then drills down into intermediate behavior.
3.Good practical relevance 3.
- 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.
1.The paper often uses language suggesting that a stage is “responsible” for disparities, but the analyses are mainly descriptive and correlational. For example, tool-exposure bias conditions on tool usage, but tool usage is itself chosen by the planner and may depend on question type, difficulty, or latent case characteristics. Likewise, transition differences do not show that those transitions caused end-to-end disparities. Without controlled interventions, causal attribution is weak.
2.Because the proposed framework still uses standard group disparity metrics and mostly audits subgroup performance at different internal stages, the conceptual jump from prior fairness auditing is not enormous. The paper should more clearly position itself as process-level auditing for agents, rather than implying a fundamentally new fairness paradigm.
3.MIMIC-FairnessVQA is useful, but the benchmark is partly constructed using LLM-generated six-choice questions from reports, followed by manual verification.
That introduces possible concerns: (1) generated questions may reflect LLM stylistic priors rather than authentic clinician questioning, (2) fairness patterns may partly come from the generated-question distribution, (3) answer verification is mentioned, but the paper gives little detail on annotation protocol, inter-rater reliability, or failure cases.
- 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
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.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The work is especially valuable for its process-level perspective and potential impact on real clinical pipelines, as well as its dataset contribution. However, the overall contribution is limited by relatively modest methodological novelty (largely organizing existing fairness ideas), a lack of causal or intervention-based validation of the identified bias sources, a narrow experimental scope focused on one agent framework, and somewhat weak proxies for reasoning bias. Overall, it is a solid and useful paper, but not strong enough in innovation or depth for a higher recommendation.
- 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
Review #2
- Please describe the contribution of the paper
The paper proposes DUCX, a stage-wise fairness audit framework for tool-using chest X-ray agents. The authors argue that fairness in agentic medical systems is not a direct extension of model-level fairness because the multi-step pipeline, such as planning, tool selection, and tool execution, creates new failure modes. They decompose end-to-end demographic disparities into three agent-specific process-level sources, which are list below:
1.Tool-exposure bias: subgroup accuracy gap conditioned on a given tool being invoked. 2.Tool-transition bias: subgroup difference in Markov transition matrices between tools. 3.LLM reasoning bias: subgroup gaps in LLM-judge-scored reasoning quality, hedging-cue frequency, and explicit demographic-term frequency in the final response.
The authors analysis 5 driver LLMs on CheXAgentBench and on a new curated benchmark, MIMIC-FairnessVQA, which contains 2000 six-choice VQA instances curated from 400 demographically balanced studies.
The paper reports that (i) end-to-end gaps persist across all five driver LLMs on both datasets (EoD up to 20.79%, FUT as low as 28.65%), and (ii) process-level analysis uncovers much larger stage-specific disparities that are invisible end-to-end, most notably a nearly 50% conditional accuracy gap when segmentation is invoked on CheXAgentBench, subgroup-asymmetric tool-routing patterns from the heatmap, and model-dependent hedging-style gaps in the final response. Together, these observations support the central claim that end-to-end fairness metrics alone cannot localize where disparity arises in an agentic pipeline.
- 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.
1.The shift from end-to-end to process-level fairness auditing for agentic system is novel and impactful. Rather than treating the agent as a black box and only measuring disparity on its final prediction, the paper argues that it is both reasonable and practically valuable to step inside the pipeline and attribute disparity to the stage where it originates. This is a well-motivated move as an end-to-end fairness number cannot tell whether a subgroup gap is inherited from a particular tool, introduced by the planner’s tool-selection and routing, or amplified during final response generation, and consequently it gives no actionable guidance about where to intervene. The paper’s stage-wise framing makes each of these pathways explicit and individually diagnosable, which is a genuine conceptual contribution beyond single-decision-function fairness audits.
2.Beyond the general motivation for process-level auditing, the paper’s specific contribution is decomposing process-level disparity into three distinct sources that correspond to the decision points of a tool-using agent, including tool-exposure bias (i.e., does a given tool deliver unequal utility when invoked for different subgroups?), tool-transition bias (i.e., does the planner route different subgroups through systematically different tool chains?), and LLM reasoning bias in the final report (i.e., does response synthesis differ by subgroup in quality, uncertainty expression, or demographic framing?). Proposing and naming this particular three-way decomposition is the novel piece. It gives each stage its own targeted diagnostic and provides the community with a usable vocabulary and a starting template for future agent-fairness work.
3.Comprehensive experimental analysis. The paper evaluates five driver LLMs, spanning open-source LLMs, VLMs, and proprietary systems, under a unified setup on two benchmarks. This breadth enables the authors to distinguish backbone-specific behaviors from stage-level patterns that remain consistent across models, which is a necessary condition for the decomposition findings to generalize beyond a single LLM. End-to-end fairness claims are further supported by non-parametric bootstrapping with reported confidence intervals in Table 1. 4.New demographic-aware benchmark. MIMIC-FairnessVQA, curated with balanced sampling over gender and age-60 threshold with five task focuses per study, partially addresses the scarcity of agent-suitable fairness datasets and may serve as a reusable resource for broader research on demographic fairness in medical visual question answering.
- 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.
1.The paper only diagnoses the problem and does not try to fix it. It describes where the unfairness comes from across the three stages, but proposes no method to reduce the disparity, and stage-targeted debiasing is only mentioned as future work in the conclusion. For a MICCAI audience, this leaves the practical impact narrow as readers learn that the agentic pipeline clearly exposes bias, but are not given a way to make the agent fairer.
2.The proposed metrics are tightly tied to MedRAX’s linear ReAct chain and do not cleanly extend to richer agent architectures. The authors justify using MedRAX because decomposition is “well-defined and modular” in a single-agent ReAct pipeline, which is an honest but narrow choice. Specifically, Delta_TTB loses its Markov interpretation in tree-based, DAG-based, or multi-agent pipelines. One can still count tool-to-tool transitions in these settings, but the resulting matrix is no longer a proper Markov transition matrix, because a node can have multiple successors, different sub-agents have their own tool chains, and P(next tool | current tool) is therefore not well-defined. The matrix then becomes a descriptive frequency table rather than a probabilistic routing policy, which weakens the claim that it captures “planning-level disparity”. Delta_TEB also becomes too coarse in multi-agent settings: the “was this tool used” indicator still applies, but when the same tool may be invoked by several sub-agents, Delta_TEB cannot tell which invocation actually caused the downstream error. The title’s broad framing, “Tool-Using Chest X-ray Agents”, therefore over-promises relative to what the three-stage decomposition structurally supports.
3.Tool-exposure bias may be confounded by case-difficulty differences across subgroups. Delta_TEB(A) is measured only on the subset of cases where tool A was invoked, but this subset is not a random sample. The driver LLM decides when to call A, and this decision depends on the image and case context. For example, if the LLM tends to call the segmenter more often on visually harder cases for one subgroup and on easier cases for the other, then the accuracy gap within “segmenter-used” cases will mix two different effects: the segmenter genuinely performing unequally across subgroups, and systematically different kinds of cases being routed to the segmenter for each subgroup. This means the headline 50% gap on CheXAgentBench cannot be read as tool bias alone. The paper reports the exposure rate P(E_A=1 | g), which tells us how often each subgroup triggers the tool but not whether the triggered cases are comparable in difficulty. A stronger analysis would match cases of similar difficulty across subgroups before computing Delta_TEB, or re-weight each case by how likely the driver LLM was to invoke the tool on it, so that cases the LLM preferred to route through the tool do not dominate the comparison.
4.Tool-transition bias is not actually reduced to a quantitative fairness evaluation. Under an ideal fair routing policy, P^(g1) = P^(g2), so Delta_TTB = 0 (the zero matrix). The paper, however, never collapses Delta_TTB into a scalar (e.g., Frobenius norm, entrywise L1, total variation distance, or row-wise KL divergence), so there is no criterion for when a transition matrix is “close enough” to the fair ideal, and no quantity that can be compared across driver LLMs, datasets, or attributes. The only evaluation is visual inspection of red/blue heatmaps (Fig. 3) together with qualitative prose observations (e.g., “female patients are more likely than male patients to proceed directly from the Classifier”). There is also no global significance test (e.g., chi-squared on transition-count contingency tables), which contrasts with the bootstrap rigor applied to end-to-end metrics in Table 1.As a result, “tool-transition bias” functions as a description, not a fairness measurement.
5.Sensitive attributes are limited to two binary axes. Only gender (M/F) and age (>=60 vs <60) are considered. Race/ethnicity is absent, despite being the most studied axis in prior CXR fairness work, and the age threshold is fixed at 60 without a sensitivity analysis.
6.Per-subgroup base rates are never reported, and the appropriateness of DP under genuine prevalence differences is not discussed. Disease prevalence often differs across demographic subgroups in medical imaging (e.g., osteoporosis by sex/age), so a model that accurately reflects such differences can still violate Demographic Parity (Chouldechova 2017[1]). Without P(y=1 | g) per subgroup on either benchmark, a reader cannot tell whether a non-zero DP in Table 1 reflects model bias or biological prevalence. Reporting per-subgroup positive-class rates, which the authors already have, and discussing when DP is informative vs. misleading would make the fairness claims more defensible.
7.Minor notation issue. Eq. 3 reuses the symbol N for the tool-pool size (T = {A_1, …, A_N} and P^(g) in R^(N x N)), which conflicts with N = dataset size defined in Sec. 2.2 (Eq. 1); the Markov matrix dimension should use |T| or a distinct symbol.
[1] Chouldechova A. Fair prediction with disparate impact: A study of bias in recidivism prediction instruments[J]. Big data, 2017, 5(2): 153-163.
- 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
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.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Overall, this paper introduces process-level auditing for fairness in agentic medical pipelines, which is a useful conceptual contribution. This reframing, together with the empirical finding that end-to-end fairness numbers can hide much larger stage-level disparities, is in my view sufficient to justify marginal acceptance at MICCAI even without a new technical primitive, because it is likely to seed follow-up mitigation work in a research direction that is becoming increasingly central. However, the paper reads primarily as a benchmark / audit paper with limited technical novelty, and I therefore rate it as a weak accept.
Key strengths: It is the first systematic process-level fairness audit of a tool-using medical agent, and the three-stage decomposition into tool-exposure, tool-transition, and LLM-reasoning bias is a useful conceptual reframing that gives the community a clear vocabulary for agent-specific disparities. The empirical findings are non-trivial: a near-50% subgroup accuracy gap conditional on segmentation-tool use, subgroup-asymmetric tool-routing patterns, and model-dependent hedging gaps, all of which are invisible from end-to-end accuracy. The evaluation spans five driver LLMs and two benchmarks with bootstrap CIs on end-to-end metrics, which is broad enough to separate backbone-specific behaviors from patterns that generalize across models.
Key limitations: It does not introduce a new fairness metric, as Delta_ACC, DP, EoD, and FUT are taken from prior works, Markov transition matrices are classical, and the reasoning-bias features (i.e., hedge-cue counting, demographic-term matching, LLM-as-judge) are also adapted from prior NLP works. It does not introduce a new agent architecture or a new training algorithm. Most importantly, it also does not propose any debiasing method. Even a lightweight mitigation at any one of the three stages (e.g., exposure rebalancing for the biased tool, a fairness-regularized prompt for tool-routing, or calibrated response synthesis) would substantially strengthen the work by turning the diagnostic framework into an actionable one for building fairer agents. Judged as a methodology paper, the contribution is genuinely incremental.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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
Review #3
- Please describe the contribution of the paper
This paper introduces DUCX, a systematic framework designed to decompose end-to-end demographic unfairness in LLM-powered, tool-using chest X-ray agents. The authors disentangle the overall bias into three distinct, quantifiable sources: tool-exposure bias, tool-transition bias, and LLM reasoning bias. This approach enables fine-grained, pipeline-level fairness auditing that goes beyond traditional black-box evaluations. Additionally, the paper contributes MIMIC-FairnessVQA, a demographic-balanced benchmark dataset tailored for evaluating chest X-ray agents. Extensive experiments across two datasets and five representative LLMs demonstrate that intermediate pipeline stages introduce and amplify biases that are often masked by end-to-end metrics.
- 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.
1.Timely and critical problem formulation: The shift from standalone models to agentic systems in medical AI introduces new pathways for bias. By addressing pipeline biases in LLM-powered tool-using agents, this work tackles a highly relevant and previously overlooked issue in clinical AI fairness. 2.Novel decomposition framework: The DUCX framework provides a clear, traceable methodology for attributing unfairness to specific stages of the agent’s execution (exposure, transition, and reasoning). This granular approach is a significant advancement over standard end-to-end fairness metrics (e.g., Equalized Odds), offering actionable insights into where disparities originate. 3.Valuable new benchmark: The curation of MIMIC-FairnessVQA, featuring balanced gender and age attributes alongside standardized multiple-choice questions, provides a much-needed resource for the community to evaluate fairness in medical VQA agents.
- 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.
1.Limited agent architectures: The evaluation is currently restricted to a single agent framework (MedRAX). While MedRAX provides the necessary trajectory observability for the DUCX decomposition, validating the framework on alternative agent architectures (e.g., multi-agent systems or different planning paradigms) would strengthen the claims of generalizability. 2.Scope of demographic attributes: The analysis focuses exclusively on gender and age (binarized at 60 years). Expanding the evaluation to include other critical attributes, such as race, ethnicity, or intersectional subgroups, would provide a more comprehensive picture of the fairness landscape in these systems. 3.Lack of clinical grounding: While the algorithmic disparities are well-documented, the paper lacks an assessment of how these specific biases (e.g., differences in hedging or tool transitions) translate into tangible clinical risks or diagnostic errors in a real-world setting. A qualitative analysis involving clinical experts would add significant value.
- 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
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.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper makes a strong and original contribution to fairness analysis for LLM-based medical imaging agents. The proposed DUCX framework provides a novel, systematic way to decompose unfairness into tool-exposure, tool-transition, and LLM reasoning bias, addressing a key gap in existing black-box fairness evaluation. The authors also contribute a new demographic-balanced dataset (MIMIC-FairnessVQA) and conduct comprehensive, well-validated experiments across two datasets and five LLMs. While the work focuses on auditing rather than mitigation and only considers gender and age, these do not weaken its core novelty, rigor, and clinical relevance.
- 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 and positive feedback and address major concerns below. A. Diagnosis without mitigation (R1, R2-W1, R3) DUCX is intentionally scoped as diagnostic, the first process-level fairness audit for tool-using medical agents. We believe rigorous diagnosis must precede principled intervention: without knowing where disparity originates, mitigation risks targeting the wrong stage. Our decomposition directly enables stage-targeted debiasing (e.g., exposure rebalancing, fairness-constrained routing) by providing the map that makes each intervention well-defined. This parallels how FairMedFM established foundation-model fairness benchmarking without proposing debiasing — a contribution valued precisely because it grounded subsequent mitigation work. B. Generalizability beyond MedRAX (R1, R2-W2, R3-W1) MedRAX was chosen for trajectory observability and modularity, prerequisites for meaningful decomposition. MedRAX is not a convenience choice but the necessary first testbed: process-level fairness decomposition requires observable trajectories and modular tool calls. In DAG or multi-agent systems, the same principle extends from Markov transitions to graph-edge or sub-agent transition statistics; this is an extension of bookkeeping, not a limitation of the core audit idea. At the same time, to the best of our knowledge, prior work has not explicitly investigated fairness evaluation and mitigation in agentic medical AI systems. We therefore chose MedRAX, a representative tool-using medical agent, as an appropriate starting point for methodological development at this early stage. The proposed framework is readily extensible, and we plan to broaden the evaluation to additional classes of medical agents in future work. C. Causal attribution and confounding (R1-W1, R2-W3, R2-W4) We note that strict causal attribution in agentic pipelines requires counterfactual trajectories, intervening on the planner’s tool choices while holding all else fixed, a problem that remains open even in the causal inference literature, where identifying causal effects in sequential decision processes with endogenous actions is an active area of research. It is therefore not a standard expectation for a fairness audit paper. That said, our design partially mitigates confounding: conditioning on tool exposure is a stratification strategy that isolates within-tool disparity from routing differences, and reporting per-group exposure rates P(E_A=1|g) makes the non-random routing visible to the reader. We present the 50% conditional gap as a signal warranting targeted investigation, not as proof of tool-level bias in isolation. For Delta_TTB (R2-W4), the matrix is intentionally primary because it identifies which routing edge shifts by subgroup; scalar summaries can complement this presentation but would obscure the actionable source of disparity. D. Limited demographic attributes (R2-W5, R3-W2) We focus on age and gender because they are consistently available across the evaluated agent benchmarks and support balanced, statistically meaningful subgroup analysis. Race, ethnicity and intersectional attributes are important, but using them responsibly requires reliable metadata and sufficient subgroup sizes; otherwise the audit risks unstable or misleading conclusions. E. MIMIC-FairnessVQA curation (R1-W3) Questions were generated using structured templates following the MedRAX-style protocol to reduce stylistic drift. Two annotators independently verified answer correctness against the original reports, and disagreements were resolved by consensus. F. DP under prevalence differences (R2-W6); Notation (R2-W7) We use DP alongside EoD because each captures different fairness desiderata. We will report per-subgroup positive-class rates and discuss when DP is informative vs. misleading under prevalence shifts, per Chouldechova (2017).
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
Consistent positive ratings are provided by reviewers, with high quality, novelty, and relevance acknowledged. The authors are encouraged to address the remaining issues to a reasonable extent in the camera-ready version.
