Machine learning for chest radiographs is shifting from reading one image at a time to comparing a patient's current scan against an earlier one, as radiologists do. The goal is to say whether disease is improving, stable, or worsening. We know that single-image classifiers underdiagnose some patient groups, such as women and young people. Nobody has checked whether comparison models inherit, reduce, or amplify those gaps. This proposal argues the question is sharper than a routine fairness audit for a concrete reason. The patients that epidemic research identifies as most vulnerable, older and male, are also the patients most likely to be imaged at the bedside, at short and irregular intervals, with a projection that differs from their earlier film. Comparison models are known to confuse changes in imaging geometry with real disease change.
Our central hypothesis is that comparison models show demographic error gaps, and that a large share of those gaps is carried by acquisition differences that track demographics: mismatched views, portable films, and short or long gaps between scans. The gaps are not mainly a matter of the model lacking knowledge of age and sex. If the hypothesis holds, the right fix is to tell the model how each image was taken, not to tell it who the patient is. Feeding demographics into the model risks teaching it base-rate shortcuts such as "old man, therefore worsening." We will measure the gaps, decompose them, and test both fixes against each other.
Why this question, and what is already known
The demographic literature supplies the motivation. A meta-analysis of 148 early COVID-19 studies, Jutzeler et al., found older age and male sex strongly associated with in-hospital mortality. A Boston admission cohort, Li et al., built integer risk scores in which male sex enters the kidney-injury score and age enters the death score. Both studies use static, point-of-admission variables. They establish who deteriorates, not how imaging models track deterioration.
The longitudinal literature supplies the models and the known weakness. Four models compare a prior and a current scan in different ways:
- DeltaNet subtracts prior-scan features from current-scan features to write COVID-19 reports.
- BioViL-T splits image features into a "current state" part and an "interval change" part without pixel registration.
- CoCa-CXR restricts attention so each region of the current scan looks only at the same region of the prior scan.
- Hu et al. build graphs over anatomical regions and compare them across time.
All four are demographic-agnostic. None reports subgroup error. The research file notes that differences between bedside front-to-back (AP) films and standing back-to-front (PA) films create pseudo-pathological feature shifts that degrade this kind of difference modeling. Our hypothesis turns that observation into a fairness mechanism.
Two works sit closest to this proposal.
- CheXclusion measured true-positive-rate gaps in static classifiers: 0.190 average sex gap on NIH, and 0.215 average age gap on its pooled dataset with ages 0–20 worst off. It also showed these gaps do not simply follow subgroup disease prevalence. It never evaluates paired scans.
- FedTAR clusters age and sex vectors and uses them to generate personalized adapter weights for multi-visit report generation. It reports CIDEr of 42.80 against 31.70 for federated averaging on J-MID. It treats demographics as a conditioning code for text quality. It does not ask whether the code helps by encoding base rates, and it does not measure subgroup error.
Our proposal differs from CheXclusion by studying comparison models and progression labels. It differs from FedTAR by treating demographic conditioning as a hypothesis to falsify, not a component to add. The specific worry is that conditioning helps aggregate scores while hurting the patients whose images contradict their group's base rate.
Core Claim
Progression models make demographically patterned errors on stable and improving cases. Older and male patients receive excess false "worsening" calls, and younger and female patients receive excess missed worsening. Most of each gap disappears when the comparison is reweighted so every group has the same distribution of view match, portable status, and inter-scan interval.
Correspondingly, giving the model acquisition information plus training on deliberately mismatched-view pairs should close more of the gap than giving it age and sex tokens. It should do so without making predictions depend on demographics when the images are held fixed.
Research Program
The first component is a stratified audit. We will join MIMIC-CXR-derived progression benchmarks to patient age and sex and to each image's acquisition metadata: view position, portable flag, and study timestamps. The benchmarks are MS-CXR-T (1,326 labeled image pairs), MMXU-test (3,000 radiologist-validated region questions), and the difference questions of MIMIC-Diff-VQA. For each model and each group, we will report three per-class rates:
- the false-worsening rate on cases whose reference label is improving or stable;
- the missed-worsening rate on cases whose reference label is worsening;
- calibration.
We use per-class rates rather than accuracy alone because accuracy hides which direction errors go.
The second component is a decomposition of each gap. Let $e_g$ be a model's error rate in group $g$ and $r$ the reference group. We estimate $\tilde e_g$, the error group $g$ would show if its pairs had group $r$'s joint distribution of acquisition covariates $a$ (view match, portable flag, binned interval). The estimate uses importance weights $w(a)=p_r(a)/p_g(a)$ fitted within each disease label so prevalence is held fixed. The share of the gap explained by acquisition is then
$$ S_g = \frac{(e_g - e_r) - (\tilde e_g - e_r)}{e_g - e_r}. $$
Our hypothesis predicts $S_g$ above one half for the older and male groups on false-worsening errors. We will also run a matched-pair analysis within patients. Many patients contribute both view-matched and view-mismatched pairs, so we can compare their error rates with patient identity held fixed.
The third component compares fixes on the same backbone, the regional cross-attention model of CoCa-CXR:
- Acquisition-aware variant. We add learned tokens for view position, portable status, and interval, in the spirit of TiBiX's interval tokens (TiBiX). We also train on synthetic "stable" pairs built from same-session or near-session films of differing views, labeled as no change.
- Demographic variant. We add age and sex tokens, matching FedTAR's design choice.
- Combined variant. Both sets of tokens.
For each variant we run a counterfactual probe: hold both images fixed, swap the demographic tokens, and count how often the predicted label flips. A model that flips often is using a prior, not the pixels.
For an epidemic-specific check, we will repeat the audit on a viral-pneumonia subset of MIMIC-CXR pairs defined from report text. We will use the DeltaNet COVID-19 cohort only if its demographic fields are available to us. The file does not say they are.
Baselines
Snapshot reference point. A DenseNet-121 single-image classifier in the CheXclusion setup predicts progression by comparing its two independent per-image probabilities. This tests whether comparison models are fairer than not comparing at all. The comparison is fair because it uses the same training pairs and the same tuning budget. "Beating it" means a smaller false-worsening and missed-worsening gap at equal or better overall accuracy. Its static sex gap of 0.190 on NIH (CheXclusion) is context only, because it is measured on a different task.
Strongest pairwise comparison model. This is CoCa-CXR, which reports 65.0% average test accuracy on MS-CXR-T progression, 4.8 points above BioViL-T (CoCa-CXR). We will retrain it from the released recipe on the CXR-4 data with three seeds. Our variants are fine-tuned from the same checkpoint with identical steps and hyperparameter sweeps, so the only difference is the extra tokens and pairs.
Strongest benchmark result on region-level progression. This is the MMXU supervised InternVL2 8B with prior medical records, reported at 88.1% on MMXU-test (MMXU). We will reproduce it and audit it as-is, since retrieval of prior records could itself carry demographic patterns. We also include the BioViL-T and Hu et al. graph models as additional audited systems. No reported subgroup-stratified result was found on Lacuna for any longitudinal model, so the disparity numbers for every baseline will be our own.
Win condition and key ablation. A variant wins if it reduces the worst-group false-worsening gap by at least a third relative to the CoCa-CXR baseline, loses no more than one point of overall MS-CXR-T accuracy, and shows a demographic-swap flip rate near zero. The ablation that isolates our key ingredient removes the mismatched-view synthetic pairs while keeping the acquisition tokens. This separates "knowing the view" from "being trained to ignore it."
Why not just…?
Why not just add age and sex as input tokens? This is the obvious move and FedTAR's design. It may well raise aggregate accuracy, because base rates differ by group. The concern is precisely that it raises accuracy by predicting the group's typical outcome: an older man with a truly stable film gets called worse. We run this variant directly. If it closes gaps and shows a low flip rate under demographic swaps, our hypothesis is wrong and we will say so. If it improves accuracy only with a high flip rate, it is buying accuracy with a prior.
Why not just rebalance or reweight subgroups during training? CheXclusion found that static disparities do not track subgroup prevalence, which already argues against prevalence balancing as a cure (CheXclusion). If our decomposition is right, demographic reweighting will leave the acquisition-driven gap untouched. We include group-reweighted training of the CoCa-CXR baseline under the same budget to confirm this.
Why not just drop view-mismatched pairs? This is the real threat. It may shrink measured gaps as well as our method does. However, mismatched pairs are concentrated in exactly the sickest, most vulnerable patients. Excluding them makes the model look fair by refusing to read the films of the people who most need reading. We will report it as a baseline alongside its per-group coverage. A filtering approach that matches our gap reduction while dropping, say, a large share of older male patients' pairs is not a solution. If its coverage loss turns out to be small and even across groups, filtering is the better recommendation and we will make it.
Why not just feed the prior report instead of the prior image? MMXU's ablation found that removing the prior image and keeping only text lowered accuracy (MMXU). Text-only comparison also inherits whatever demographic patterns exist in report writing. We audit the text-prior configuration as part of the MMXU baseline, so this is measured rather than assumed.
Proposed Experiments
The decisive experiment is the decomposition on MS-CXR-T and MMXU-test. If acquisition reweighting explains less than a third of the age and sex gaps for every model, the mechanism is wrong. In that case the proposal reduces to a first audit of comparison models, which is still publishable, and the fix comparison becomes moot.
The second check is the within-patient matched-pair contrast. Error should rise on view-mismatched pairs relative to matched pairs for the same patient, and the rise should be larger in older patients if acquisition severity also differs by age.
The third check is the counterfactual demographic swap across all three variants.
The fourth check is transfer. Variants trained on CXR-4 are evaluated on MIMIC-Diff-VQA difference questions, where Hu et al. report a CIDEr of 1.027 (Hu et al.). Here we test whether gap reductions hold on a benchmark the fixes were not tuned on.
All results will be reported with patient-level bootstrap intervals, because a single patient contributes many pairs and treating pairs as independent would overstate certainty.
References
- Comorbidities, clinical signs and symptoms, laboratory findings, imaging features, treatment strategies, and outcomes in adult and pediatric patients with COVID-19: A systematic review and meta-analysis
- Risk assessment for acute kidney injury and death among new COVID-19 positive adult patients without chronic kidney disease: retrospective cohort study among three US hospitals
- DeltaNet: Conditional Medical Report Generation for COVID-19 Diagnosis
- Learning to Exploit Temporal Structure for Biomedical Vision-Language Processing
- CoCa-CXR: Contrastive Captioners Learn Strong Temporal Structures for Chest X-Ray Vision-Language Understanding
- Expert Knowledge-Aware Image Difference Graph Representation Learning for Difference-Aware Medical Visual Question Answering
- CheXclusion: Fairness gaps in deep chest X-ray classifiers
- Personalized Longitudinal Medical Report Generation via Temporally-Aware Federated Adaptation
- TiBiX: Leveraging Temporal Information for Bidirectional X-Ray and Report Generation
- MMXU: A Multi-Modal and Multi-X-ray Understanding Dataset for Disease Progression