{"id":"5804c3e2-02e2-440d-b199-ae25c1638810","arxiv_id":"2504.19621","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"CIT-LR, a diffusion-based statistical test, checks whether medical imaging classifiers change their predictions when a patient's reported race is altered in generated counterfactual images.","lead":"The paper builds a test that generates altered chest X-rays with a different reported race, then checks whether an AI diagnosis changes; stable predictions are called counterfactually invariant. It aims to catch hidden race bias in medical imaging AI without needing real images of the same patient with a different race.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Alg. 1's t-statistic is mathematically invalid, so the reported p-values and the claimed ECA alignment are unsupported as written.","rationale":"I agree with the reader's rejection, but I would anchor it on an even more basic failing: the test statistic in Algorithm 1 is not a valid t-statistic, so the p-values cannot be interpreted even before considering the adjustment-set assumption. The reader's weakest_assumption field focused on the latent representation being a valid adjustment set, which is a real and self-admitted limitation, but the t-statistic bug is decisive and easier to verify from the paper itself. The reader's rationale did mention the invalid t-test, which is why my agreement is partial rather than full. A corrected t-statistic might rescue the method in principle, and the circular ECA validation could be addressed by an external counterfactual benchmark, but none of that is present in the manuscript. The verdict of REJECT stands, with no change from the reader's assessment.","tokens_in":37305,"tokens_out":7215,"duration_ms":80011,"concrete_test":"Use the public code to extract the per-sample values u_i = yhat_i * ghat(a_i, z_i) and v_i = yhat_i * hhat(z_i) for the synthetic, cheXpert, and MIMIC-CXR runs; recompute p-values with the correct paired t-test t = mean(u-v) / (sd(u-v)/sqrt(n)) and also with a Welch two-sample test; compare the reject/accept decisions at alpha=0.05 to those in Figs. 2-4. If any panel changes the set of rejected classifiers, the headline claim about consistency with ECA is not supported by the procedure as written. If the code is unavailable, reimplementing Alg. 1 lines 6-16 from the paper suffices to test the formula.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is that Algorithm 1 does not actually implement a valid statistical test. Line 16 computes t = sqrt(|Dtest|)(mu_Z - mu_AZ)/(sigma_Z - sigma_AZ) and refers to a two-sample Student's t-test with |Dtest|-1 degrees of freedom. This is not the paired t-statistic, whose denominator is the standard deviation of the per-sample differences divided by sqrt(n), nor the two-sample t-statistic, whose denominator is the standard error of the difference of means. The denominator here is the difference of two estimated dispersions, which can be zero or negative, making the statistic unstable and uninterpretable. Every p-value reported in Figs. 2-4, and hence the central empirical claim that CIT-LR 'rejects when ECA is low and does not when the ECA is high,' rests on this invalid statistic. A second, compounding problem is that the real-world ECA ground truth is computed from the same CLDM and the same adjustment-set assumption used by the test, so agreement with ECA on cheXpert/MIMIC-CXR is not an independent confirmation. The paper itself concedes in the Impact Statement that the key representation assumption is not adequately proved. The t-statistic error alone is sufficient to invalidate the reported results as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CIT-LR, a statistical test for counterfactual invariance (CI) of a pretrained medical imaging classifier with respect to a sensitive attribute A. The test combines a conditional latent diffusion model (CLDM) with a disentangling transformation to learn the functions g(a,z)=E[Y|A=a,Z=z] and h(z)=E[Y|Z=z], then uses a t-test on E[Y g(A,Z)] versus E[Y h(Z)] to decide CI. Experiments on synthetic data, cheXpert, and MIMIC-CXR compare CIT-LR p-values with an Expected Counterfactual Accuracy (ECA) ground truth and with DP/EO baselines, claiming stronger alignment. As written, however, the central statistical step is invalid and the validation is partly circular.","tokens_in":37522,"tokens_out":5702,"duration_ms":57596,"significance":"The intended contribution—a practical, falsifiable statistical audit for demographic bias in diagnostic models—is important and timely. If the test were valid, it would improve on association-based fairness metrics by targeting counterfactual invariance. The paper also makes useful practical steps: it provides code, considers 100 classifiers per dataset, and reports comparisons with DP/EO. However, the significance is currently not realized: the t-statistic in Algorithm 1 is mathematically incorrect, the quantities computed in Step 2 do not match the functions learned in Step 1, and the real-world ECA ground truth is derived from the same unproved adjustment-set assumption that the test uses. These issues undermine the central empirical claim.","major_comments":[{"comment":"The t-statistic t = sqrt(|Dtest|)(mu_Z - mu_AZ)/(sigma_Z - sigma_AZ) with |Dtest|-1 degrees of freedom is not a valid test statistic for any standard hypothesis test. The denominator is the difference of two estimated standard deviations, not a standard error of the difference of means; it can be zero or negative, making the statistic unstable and uninterpretable. The surrounding text says a paired t-test is used, while the algorithm says a two-sample Student's t-test; neither would have this form. Since every p-value reported in Figures 2-4 and Tables 12-13 is computed from this invalid statistic, the central empirical claim that CIT-LR 'rejects when ECA is low and does not when the ECA is high' is unsupported as written.","section":"Section 5, Algorithm 1, Step 3 (line 16)"},{"comment":"There is a mismatch between the functions learned in Step 1 and the inputs on which they are evaluated in Step 2. Step 1 defines ghat(a,z) and hhat(z) using zhat_0 ~ p_theta(zhat_0 | z,a), i.e., reverse-diffusion samples from the CLDM. Step 2, however, computes ghat(a, ztilde_T) and hhat(ztilde_T) with ztilde_T ~ q_theta(ztilde_T | ztilde_0), i.e., forward-diffusion noised latents. These are different random variables, so the test is not actually evaluating E[Y|A=a,Z=z] or E[Y|Z=z] at the representation used in Theorem 4.2. The estimated quantities therefore do not correspond to the null hypothesis of the test.","section":"Section 5, Algorithm 1, Steps 1-2"},{"comment":"The real-world ECA ground truth is not independent confirmation of the method. In Section 6.3 and Appendix E, ECA is computed using the equality E[Y|do(A=a),Z=z] = E[Y|A=a,Z=z] from Lemma D.3, with Z being the learned CLDM latent representation. The test's null hypothesis in Theorem 4.2 relies on exactly the same equality and the same Z. Hence the alignment between CIT-LR p-values and ECA on cheXpert and MIMIC-CXR is partly by construction and does not provide independent evidence that the test detects true counterfactual invariance. The paper's Impact Statement concedes that the key assumption that latent representations capture all relevant factors 'is not adequately proved.'","section":"Section 6.3 and Appendix E"},{"comment":"The disentanglement condition Z perpendicular A is not established. The optimization minimizes a lower bound on the mutual information I(phi(z0), a) using the MINE estimator, not the true mutual information, and no theorem guarantees that the resulting representation satisfies the valid-adjustment-set condition required by Theorem 4.2 and Lemma D.3. This is a load-bearing gap in the identification argument, since the validity of the entire test depends on this assumption.","section":"Section 3.3, Eq. (3)"}],"minor_comments":[{"comment":"The terminology is inconsistent: the text says a paired t-test is used, while Algorithm 1 Step 3 says a two-sample Student's t-test. The paper should state one correct test and use it consistently.","section":"Section 5"},{"comment":"There are numerous typos and formatting issues, including 'spECAfic' in place of 'specific', 'ff.' in place of 'if and only if', 'DECAsionTreeClassifier' in Table 1, and 'Equality of Opportunity' versus 'Equal Opportunity' in Section 6.1.","section":"Throughout"},{"comment":"The figures plot log-transformed p-values, but the handling of p-values that are exactly zero or near machine precision is not described; clarify the transformation used.","section":"Figures 2-4"},{"comment":"The reference list contains several duplicate entries (e.g., Dhariwal and Nichol 2021a/2021b, Müller-Franzes et al. 2023a/2023b) and some incomplete entries; please consolidate and complete the bibliography.","section":"References"},{"comment":"The baseline tests for DP and EO are described as paired t-tests, but the exact test statistic and degrees of freedom are not specified, making the comparison with CIT-LR difficult to assess even setting aside the invalid statistic in Algorithm 1.","section":"Section 6.1"}],"recommendation":"reject","confidential_remarks":"The paper addresses an important problem and includes a substantial experimental effort, but the central statistical test is mathematically invalid and the real-world validation is circular. A corrected version would need to fix the t-statistic, align the learned functions with the evaluated quantities, and provide an independent ground truth for counterfactual invariance, which goes beyond a minor revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline claim is not supported as written. The t-statistic in Algorithm 1, line 16, is t = sqrt(|Dtest|)(mu_Z - mu_AZ)/(sigma_Z - sigma_AZ). That is not a paired t-statistic (whose denominator is the standard error of paired differences) nor a two-sample t-statistic (whose denominator is the standard error of the difference of means). The denominator here is a difference of estimated standard deviations, which can be zero or negative, making the statistic unstable and uninterpretable. Every p-value in Figs. 2–4 rests on this invalid statistic, so the central empirical claim—that CIT-LR rejects when ECA is low and does not when ECA is high—is unsupported.\n\nWhat is genuinely new is the engineering combination: disentangled conditional latent diffusion as the generator, plus a statistical test for counterfactual invariance of a pretrained classifier. On the synthetic datasets, with the DGP fully known, the idea of comparing your test's decisions against ECA is sound even if the t-statistic needs correction. The paper also ships code and gives detailed appendices on architectures and training, which is helpful and measurable.\n\nThere is a second, compounding problem: on cheXpert and MIMIC-CXR, the ECA ground truth is computed using the same CLDM-based adjustment set, E[Y|do(A=a),Z=z] = E[Y|A=a,Z=z] from Lemma D.3. The test's null hypothesis is exactly the failure of that equality, so agreement between CIT-LR p-values and ECA on real data is not independent confirmation. The authors themselves concede in the Impact Statement that the key representation assumption \"is not adequately proved.\" That is a load-bearing limitation, not a minor caveat.\n\nI also found the paper weaker on clarity: the text says both \"paired\" and \"two-sample\" t-test, and the notation sigma for both standard deviations and variances is messy. Those are minor relative to the main statistical issue.\n\nWho is this for? A fair-minded reader interested in practical counterfactual bias auditing in medical imaging will find the setup plausible and the ambition worth attention, but they should not take the reported numeric results at face value. The paper deserves a serious referee: the core problem is important, the authors have not faked anything, and the errors are distinctly fixable—recompute the test statistic properly, and design an independent ground truth, at least on synthetic data where ECA is known by construction. If those are addressed, this could become a useful tool. As is, I would not cite it yet, but I would bring it to a reading group to discuss where the method needs to change.\n\nRecommendation: send to peer review with the expectation of major revision.","headline":"The central algorithm's t-statistic is invalid as written, so the reported p-values and the claimed alignment with ECA do not hold up; still, the paper tackles a real problem and deserves a serious referee with an invitation to fix the statistics and break the circular validation.","tokens_in":38098,"tokens_out":1185,"would_cite":false,"duration_ms":14681,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A statistical test using generated counterfactual X-rays can reveal whether medical AI models are biased by sensitive attributes like race.","keywords":["counterfactual invariance","medical imaging fairness","latent diffusion models","statistical hypothesis testing","causal bias auditing","chest X-ray","demographic parity","equality of opportunity"],"falsifier":"Construct a synthetic structural causal model where the latent representation deliberately omits a disease-relevant confounder, train classifiers with known counterfactual bias, and check whether CIT-LR still tracks the true expected counterfactual accuracy; if the test's p-values stop correlating with ground truth when the adjustment set is incomplete, the key assumption is thereby refuted.","tokens_in":37015,"feed_emoji":"🩻","tokens_out":6760,"duration_ms":58548,"temperature":0.7,"pith_summary":"Medical imaging models sometimes rely on patient demographics in ways that can distort diagnoses, but standard fairness metrics can miss this. This paper tries to establish that a statistical test called CIT-LR can decide whether a diagnostic classifier is counterfactually invariant to a sensitive attribute—whether its predictions would stay the same if that attribute were changed while everything else stayed fixed. The test avoids needing real counterfactual images: it trains a conditional latent diffusion model to generate race-perturbed X-rays and then runs a hypothesis test comparing two quantities that Theorem 4.2 says are equal exactly when the model is invariant. On synthetic data and on the cheXpert and MIMIC-CXR chest X-ray datasets, the resulting p-values track a ground-truth measure of counterfactual bias more consistently than demographic parity or equality of opportunity. If the approach holds, it gives auditors a practical, quantitative tool to check pretrained diagnostic models for hidden demographic biases before deployment.","feed_headline":"Race-swapped X-rays reveal when medical AI hides demographic bias","feed_subtitle":"CIT-LR tracks counterfactual bias better than DP and EO on cheXpert and MIMIC-CXR.","key_machinery":"The load-bearing identity is Theorem 4.2: given a latent representation $Z$ that is a valid adjustment set, a classifier is counterfactually invariant iff $\\mathbb{E}[\\hat Y g(A,Z)] = \\mathbb{E}[\\hat Y h(Z)]$, which follows from a law-of-total-variance argument. The machinery that turns this identity into a test is a conditional latent diffusion model with a disentangling transform $\\phi_{\\lambda,\\delta}$ that makes the representation approximately independent of the attribute $A$ (trained by minimizing a mutual-information lower bound), used to generate counterfactual images; the test statistic is a paired $t$-test comparing the two expectations estimated on separately split datasets.","core_discovery":"The paper's central claim is that a model's counterfactual invariance with respect to a sensitive attribute can be tested from observed samples by checking the identity $\\mathbb{E}[\\hat Y g(A,Z)] = \\mathbb{E}[\\hat Y h(Z)]$, where $Z$ is a latent representation, $g(a,z) = \\mathbb{E}[\\hat Y \\mid A=a, Z=z]$ and $h(z) = \\mathbb{E}[\\hat Y \\mid Z=z]$. Under the conditional ignorability assumption that $Z$ is a valid adjustment set, this identity holds exactly when the classifier is counterfactually invariant. The proposed CIT-LR algorithm implements the test by training a disentangled conditional latent diffusion model to generate counterfactual images with the sensitive attribute changed, estimating $g$ and $h$ from those generations, and running a two-sample Student's $t$-test on $\\hat Y g(A,Z)$ versus $\\hat Y h(Z)$. The authors report that in experiments on synthetic data and on cheXpert and MIMIC-CXR, CIT-LR rejects the null hypothesis when the expected counterfactual accuracy is low and fails to reject when it is high, providing stronger consistency than demographic parity and equality of opportunity.","pith_inferences":["If the latent representation captures all disease-relevant factors, the same machinery could audit other sensitive attributes (age, sex) and other medical modalities, not just chest X-rays.","The test's validity could be checked by a 'negative control' experiment: train classifiers with deliberately injected known confounding, ablate the disease-relevant latent factor, and see whether CIT-LR's verdict flips; if it does not, the causal assumption is doing the work.","The disentangling transform $\\phi$ is optimized via a mutual-information lower bound; since lower bounds can be loose, the actual degree of disentanglement could be reported as a diagnostic of the test's reliability.","A practitioner might use CIT-LR not only to reject bad checkpoints but also to rank checkpoints of the same architecture, choosing the one whose p-value indicates invariance, before clinical deployment."],"forward_implications":["Pretrained diagnostic classifiers can be audited for sensitive-attribute bias without collecting any counterfactual images, using only the training data and a generative model.","The test splits diseases into those where near-invariant classifiers exist (consolidation) and those where most classifiers are non-invariant (cardiomegaly, edema, pleural effusion), giving a concrete target for fairness improvement.","Because it is a hypothesis test with p-values, the method provides a significance threshold for model selection or regulatory compliance rather than a point estimate.","On the tested datasets, demographic parity and equality of opportunity both flag invariant models as biased and miss biased ones; CIT-LR is claimed to be more consistent, so replacing association-based audits with a counterfactual test could change which models get deployed."],"supporting_citations":[{"why":"supplies the definition of counterfactual fairness that Def. 4.1 extends to the expectation of the classifier output.","marker":"[Kusner et al., 2017b]"},{"why":"gives the mutual-information neural lower bound used to train the disentangling transform $\\phi$.","marker":"[Belghazi et al., 2018]"},{"why":"introduces latent diffusion models that the conditional latent diffusion model in Alg. 1 is built on.","marker":"[Rombach et al., 2022]"},{"why":"provides the cheXpert chest X-ray dataset used to evaluate CIT-LR against expected counterfactual accuracy.","marker":"[Irvin et al., 2019]"},{"why":"provides the MIMIC-CXR chest X-ray dataset used as the second real-world evaluation.","marker":"[Johnson et al., 2019a,b]"},{"why":"supplies the general conditional counterfactual invariance notion and the proof strategy that Theorem 4.2 follows.","marker":"[Quinzan et al., 2024]"}],"fun_headline_variants":["Race-swapped X-rays unmask medical AI's hidden bias","Synthetic race-swaps detect unfairness in diagnostic AI","Diffusion-based counterfactuals flag biased medical AI","Counterfactual invariance test catches biased medical AI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The test collapses if the latent representation learned by the generative model does not capture every factor that matters for the diagnosis, because then the statistical equality used to judge invariance is not guaranteed to hold.","fun_headline_variants_meta":{"raw":{"variants":["Race-swapped X-rays unmask medical AI's hidden bias","Synthetic race-swaps detect unfairness in diagnostic AI","Diffusion-based counterfactuals flag biased medical AI","Counterfactual invariance test catches biased medical AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000793,"raw_usage":{"total_tokens":3519,"prompt_tokens":998,"completion_tokens":2521,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":614,"completion_tokens_details":{"reasoning_tokens":2456}},"tokens_in":614,"tokens_out":2521,"duration_ms":19105,"temperature":1.0,"reasoning_tokens":2456,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:47:25.207440+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a synthetic structural causal model where the latent representation deliberately omits a disease-relevant confounder, train classifiers with known counterfactual bias, and check whether CIT-LR still tracks the true expected counterfactual accuracy; if the test's p-values stop correlating with ground truth when the adjustment set is incomplete, the key assumption is thereby refuted.","supporting_citations":[],"review_version":1}