PaperlayerView exact revision on arXiv

Source-linked AI summary

Responsible Quantum Machine Learning for Medical Imaging: A Fairness Audit Across Sensitive Subgroups

Giovanni dos Santos Franco, Felipe Mahlow, Ellison Fernando Cardoso, José Adenaldo Santos Bittencourt, Felipe Fanchini

arXiv:2610.37819v1cs.LGcs.AIcs.CY

TL;DR

Medical AI can behave unequally across sensitive subgroups, while fairness in QML medical imaging remains underexplored. This study audits classical and quantum classifiers with mitigation strategies and finds that fairness and performance trade-offs vary by model and dataset.

  • Problem

    Medical AI can show unequal performance across sensitive subgroups, while QML classifiers remain understudied under subgroup-aware fairness criteria.

    Source: §1 · p. 2

  • Method

    The study audits classical and quantum classifiers on DDI and CheXpert-derived data, using ResNet-18 embeddings, fairness mitigation, and subgroup-aware evaluation.

    Source: §2.2 · p. 8, §2.5 · p. 12, §5 · p. 30

  • Results

    Fairness and performance trade-offs varied by model and dataset; on DDI, QCNN with GRL had an AUC gap of 0.011 while losing 0.042 AUC points relative to its baseline.

    Source: §5 · p. 31

  • Takeaways & Limitations

    A reduced fairness gap is not intrinsically beneficial; the study evaluates disparity reductions alongside diagnostic performance and non-inferiority.

    Source: §5 · p. 31

    The design cannot cleanly attribute observed fairness behaviour to the quantum component because the models use classical ResNet-18 embeddings compressed to ten principal components.

    Source: §4 · pp. 28–29

Abstract

from arXiv · show

Quantum machine learning (QML) is increasingly explored for medical classification, yet its behaviour under subgroup-aware fairness criteria remains understudied. We present an empirical fairness audit of classical and quantum classifiers for medical imaging, assuming QML is not inherently fair. Two scenarios are evaluated: dermatological lesions on DDI (sensitive attribute: Fitzpatrick skin type) and chest X-rays on CheXpert-derived data (sensitive attribute: age group). Using ResNet-18 embeddings, we train SVM, MLP, QSVM and QCNN under a baseline and three mitigation settings: differentiable Equalized Odds (EO) regularisation, Gradient Reversal Learning (GRL), and both combined. Foundation models (PanDerm, EVA-X) serve as external references. Quantum classifiers may achieve competitive aggregate performance while still exhibiting subgroup disparities. Since a smaller subgroup AUC gap may stem from degrading the best group rather than improving the worst, each mitigation is also tested against a pre-specified non-inferiority margin relative to its baseline. Under this criterion, several configurations with the smallest disparity, notably QCNN with GRL on DDI, are not clinically acceptable due to loss of global discrimination. On DDI, QSVM with GRL+EO performed best and approached PanDerm in the performance/fairness trade-off; on CheXpert, QCNN with GRL achieved the highest aggregate performance but increased disparity, and EVA-X remained stronger. Fairness effects are model- and dataset-dependent, so QML requires empirical subgroup-aware evaluation. All quantum models run in classical simulation; we claim no quantum advantage.

1 Introduction

Medical AI can show unequal performance across demographic and clinically relevant subgroups, so aggregate metrics alone do not capture how systems behave across groups. This work audits classical and quantum classifiers under fairness mitigation and evaluates disparity reductions alongside predictive performance.

1 Introduction

  • Medical AI can exhibit unequal behaviour across subgroups, including differences associated with age, sex, ethnicity, and skin tone, despite strong aggregate performance.Prior work reports limitations across Fitzpatrick skin type groups in DDI and fairness gaps and underdiagnosis bias in chest radiography datasets.Source: §1 · p. 2
  • QML studies have largely focused on predictive performance, compactness, or near-term feasibility, while fairness across sensitive groups is also relevant to healthcare-oriented applications.The introduction names QSVMs, variational quantum classifiers, and QCNNs as explored QML approaches.Source: §1 · p. 2
  • A smaller group-wise AUC gap can reflect either improvement in the worst-performing group or deterioration in the best-performing group, which are not clinically equivalent.The study therefore interprets disparity changes jointly with predictive performance relative to the unmitigated baseline, using a non-inferiority criterion.Source: §1 · pp. 2–3
  • The study contributes an empirical QML fairness audit, tests GRL and combined GRL+EO mitigation, compares classical and quantum classifiers, and introduces a non-inferiority criterion for performance–fairness trade-offs.A mitigation counts as favourable only when it reduces subgroup disparity without degrading global predictive performance beyond a pre-specified margin relative to its baseline.Source: §1 · p. 4

2.1 Study design

The study audits classical and quantum classifiers in two medical imaging scenarios, using sensitive subgroups defined by skin type or age. It also describes the datasets’ differing class distributions and repeats experiments across random seeds.

2.1 Study design

  • The design evaluates dermatological classification on DDI by Fitzpatrick skin type and CheXpert-derived pleural-effusion classification by age, comparing classical and quantum models under fairness-mitigation settings.Images are represented by ResNet-18 embeddings; PanDerm and EVA-X provide external references for the dermatology and chest X-ray settings, respectively.Source: §2.1 · p. 5
  • The experiments use 60/20/20 training, validation, and test splits and five random seeds; reported deviations are standard deviations, not resampling-based confidence intervals or paired-model estimates.They indicate configuration stability across repetitions but do not by themselves establish superiority or equivalence when intervals overlap.Source: §2.1 · p. 5

2.1.1 DDI

  • Malignant cases represented 23.6% of FST I–II, 30.7% of FST III–IV, and 23.2% of FST V–VI in DDI.The six Fitzpatrick categories were grouped into three broader groups for subgroup evaluation.Source: §2.1.1 · p. 6

2.1.2 CheXpert

  • Pleural-effusion prevalence increased across age groups, from 28.5% in ages 0–40 to 38.4% in ages 41–60 and 45.6% in ages 60+.The study uses age-aware evaluation to assess whether model behaviour varies across those groups.Source: §2.1.2 · p. 7

2.2 Feature extraction

The classifiers use fixed-dimensional embeddings extracted from ImageNet-pretrained ResNet-18. Quantum-model inputs are further compressed to ten dimensions for compatibility with the evaluated circuits.

2.2 Feature extraction

  • Both classical and quantum classifiers use ResNet-18 embeddings extracted after removing the pretrained model’s final classification layer.The shared representation reduces the possibility that model-family performance differences arise from different image encoders.Source: §2.2 · p. 8
  • Quantum-model inputs are ten-dimensional PCA projections of ResNet-18 embeddings, matching the ten-qubit circuits with one feature per qubit.PCA is used to make the representations compatible with the quantum circuits, not as a fairness intervention.Source: §2.2 · p. 8
  • The quantum models inherit representations already filtered by ResNet-18 and compressed by PCA, so subgroup bias may be established upstream of the quantum component.The design therefore places substantial representational burden, and potentially subgroup bias, in the classical stage.Source: §2.2 · p. 9

2.3 Classifiers

The study compares classical SVM and MLP baselines with quantum-kernel QSVM and variational QCNN classifiers. Their input dimensions and computational costs differ, so cross-family results describe complete pipelines rather than isolated classifier mechanisms.

2.3.1 Classical and foundation baselines

  • SVM and MLP provide classical baselines, while PanDerm and EVA-X serve as external references for dermatology and chest X-ray classification.SVM uses an RBF kernel; MLP supplies a nonlinear neural baseline distinct from the SVM.Source: §2.3.1 · p. 9
  • Comparisons between classical and quantum models reflect complete pipelines because input dimensionality, PCA, model capacity, regularisation, and quantum feature maps are not isolated.A matched-capacity classical control using the same ten-dimensional PCA inputs is identified as necessary to isolate the quantum component.Source: §2.3.1 · p. 9

2.3.2 Quantum support vector machine

  • QSVM builds a kernel matrix from fidelities between quantum-encoded sample states, with pairwise kernel evaluations scaling as O(N^2) for N training samples.Its ten-dimensional inputs are encoded with angle rotations and a feature map using a CNOT ring and feature-dependent rotations.Source: §2.3.2 · p. 10

2.3.3 Quantum convolutional neural network

  • QCNN uses ten-dimensional inputs in a trainable variational circuit with convolution-like and pooling-like blocks to learn a nonlinear decision function.Unlike QSVM, its computational cost is dominated by repeated circuit evaluations during training and depends on samples, parameters, optimisation steps, and seeds.Source: §2.3.3 · pp. 10–11

2.4 Simulation protocol and reproducibility

The quantum models were evaluated in classical simulation, with repeated train/validation/test splits and results averaged across five seeds. The reported standard deviations describe variability across runs, not resampling-based confidence intervals.

2.4 Simulation protocol and reproducibility

  • Both QSVM and QCNN used ten qubits, but the QSVM used a fidelity-kernel feature map while the QCNN was a trainable variational classifier.The QSVM's final classifier was a kernel-based SVM.Source: §2.4 · p. 11
  • All quantum computations used classical statevector simulation; no quantum hardware experiment or quantum-advantage claim was made.The audit concerns the hybrid classical–quantum pipeline as a whole.Source: §2.4 · p. 11
  • Trainable neural and variational components used Adam with learning rate 10−3, weight decay 10−4, at most 25 epochs, and early-stopping patience of 5.Source: §2.4 · p. 11
  • Experiments used five seeds and independent 60/20/20 train/validation/test splits; reported test-set values were averaged across seeds with standard deviations.The standard deviations reflect run-to-run variability, not resampling-based confidence intervals.Source: §2.4 · p. 11

2.5 Fairness mitigation strategies

The evaluation compares baseline, EO, GRL, and combined GRL+EO settings. EO regularises class-conditional group differences in predicted probabilities, while GRL discourages sensitive-group information in learned representations.

2.5 Fairness mitigation strategies

  • Four settings were evaluated: baseline, differentiable EO regularisation, GRL, and combined GRL+EO.The baseline had no explicit fairness intervention and served as the reference.Source: §2.5 · pp. 11–12
  • EO adds a penalty equal to the summed variance of group-wise mean predicted probabilities within each true class.The penalty is added to binary cross-entropy, encouraging reduced class-conditional group disparities.Source: §2.5 · pp. 12–13
  • GRL trains a shared representation for clinical classification and sensitive-group prediction, reversing the adversarial gradient to discourage sensitive information.The reversal coefficient is zero for the first five epochs, then rises linearly from zero to 1.00 through epoch 25; the adversarial loss weight is 0.15.Source: §2.5 · pp. 13–14
  • GRL treats sensitive information as undesirable even when it may be clinically relevant; in CheXpert, age also predicts pleural-effusion prevalence.This motivates testing mitigated models against a non-inferiority margin on global discrimination.Source: §2.5 · p. 13
  • GRL+EO combines representation-level adversarial training with prediction-level EO regularisation; the study does not assume both improve every fairness metric.The combined setting uses αadv = 0.15 and λEO = 0.50.Source: §2.5 · p. 14
  • For QSVM mitigation, a differentiable projection maps ResNet-18 embeddings to ten-dimensional inputs before the fidelity kernel is computed.Mitigation changes the representation used for the kernel rather than the SVM optimisation directly.Source: §2.5 · p. 14

2.6 Evaluation metrics

The evaluation combines predictive metrics with subgroup-disparity measures and judges mitigation against both disparity reduction and preservation of global discrimination. The reported metrics do not fully characterize clinical consequences at operating points.

2.6 Evaluation metrics

  • Predictive performance was measured with AUC, balanced accuracy, and F1-score, capturing ranking quality, class-balanced performance, and precision–recall balance.Source: §2.6 · p. 14
  • The main fairness metric was the group-wise AUC gap, which compares subgroup discrimination without relying on a fixed decision threshold.The gap alone does not reveal whether disparity changed through improvement of the worst-performing group or deterioration of the best-performing group.Source: §2.6 · p. 15
  • A mitigation is favourable only if it reduces subgroup disparity and keeps global discrimination within the pre-specified non-inferiority margin of 0.02 AUC points.The margin is a methodological convention, not a validated clinical threshold, and would need task- and population-specific establishment for deployment.Source: §2.6 · p. 15
  • The joint criterion avoids accepting a near-chance model with uniformly poor subgroup performance, but constrains global discrimination because subgroup AUCs were computed only for baseline configurations.A stronger criterion preventing any subgroup from deteriorating beyond ε is identified as future work.Source: §2.6 · p. 15
  • Demographic parity gap was treated as supplementary because differences in true clinical prevalence can affect it.Source: §2.6 · pp. 15–16
  • The metrics do not weight error types by clinical consequences; fuller characterization would require subgroup sensitivity, specificity, predictive values, and decision-curve analysis.AUC and AUC gap are threshold-independent; the other reported metrics use a single calibrated operating point.Source: §2.6 · p. 16

3 Results

Across DDI and CheXpert, the results examine subgroup disparities and whether fairness-mitigation strategies reduce them while preserving global discrimination. In CheXpert, no evaluated strategy brought subgroup performance to a level supporting responsible deployment, and the global-AUC non-inferiority analysis could not distinguish gains for the worst-performing group from deterioration for the best-performing group.

3.1 Global predictive performance

  • On CheXpert, QCNN with GRL had the highest non-foundation point estimates; on DDI, QSVM with GRL+EO’s estimates were comparable to PanDerm’s across AUC, balanced accuracy, and F1-score.On CheXpert, QCNN with GRL scored AUC 0.820±0.029, balanced accuracy 0.771±0.030, and F1-score 0.662±0.044. On DDI, QSVM with GRL+EO scored 0.808±0.051, 0.770±0.046, and 0.651±0.070, versus PanDerm’s 0.814, 0.766, and 0.640; no paired statistical test or formal equivalence assessment was conducted.Source: §3.1 · pp. 17–18

3.2 Subgroup ROC analysis

  • On DDI, baseline subgroup AUCs varied across Fitzpatrick groups: QSVM ranged from 0.671 to 0.900, while PanDerm’s best-to-worst gap was 0.099.The classical baselines also showed lower AUC for FST III–IV, while QCNN had lower discrimination for FST V–VI. These subgroup estimates came from a modest test set and lacked resampling-based confidence intervals, so they indicate approximate disparity rather than precise magnitudes.Source: Figure 4 · p. 19, §3.2 · pp. 19–20
  • On CheXpert, downstream models had weaker baseline discrimination for ages 0–40, with AUCs of 0.524–0.648, whereas EVA-X ranged from 0.907 to 0.945 across age groups.The youngest group had the fewest positive cases (pleural-effusion prevalence 28.5%), which may contribute to the observed weakness and points toward data-level remedies; its subgroup estimates were also based on the smallest subgroup and lacked resampling-based confidence intervals.Source: Figure 5 · p. 20, §3.2 · pp. 20–21

3.3 AUC-based performance–fairness trade-off

  • The AUC-gap trade-off favored QSVM with GRL+EO on DDI, while CheXpert’s strongest non-foundation aggregate performer, QCNN with GRL, increased disparity and failed the disparity-reduction condition.QSVM with GRL+EO had mean AUC 0.808 and AUC gap 0.083, compared with PanDerm’s 0.814 and 0.099; relative to its baseline, it reduced the gap by 0.146 and increased AUC by 0.015. QCNN with GRL on CheXpert gained 0.030 AUC points but increased its AUC gap by 0.027. Mitigation effects were model- and dataset-dependent.Source: §3.3 · pp. 21–23, Figure 6 · p. 22

3.4 AUC gap analysis

  • On DDI, QCNN with GRL and SVM with GRL had the smallest AUC gaps (0.011 and 0.020), but QCNN’s AUC fell from 0.737 to 0.695, exceeding the non-inferiority margin.On CheXpert, EVA-X had the lowest AUC gap (0.038); the lowest downstream gaps were 0.183 for MLP with Equalized Odds, 0.202 for QCNN with GRL+EO, and 0.212 for QSVM with GRL+EO.Source: Figure 7 · p. 23, §3.4 · pp. 23–24

3.5 Demographic parity gap analysis

  • Demographic-parity results differed across datasets: DDI’s lowest gaps included 0.027 for QSVM with GRL, whereas CheXpert’s lowest downstream gaps were 0.160–0.165 and remained non-negligible.Demographic parity measures differences in positive prediction rates, not group-wise discrimination; on CheXpert, age-related prevalence differences mean a residual gap may be expected and should not simply be minimized. In DDI, PanDerm had a higher demographic-parity gap than some downstream models despite strong aggregate performance, and the lowest-gap settings did not necessarily align with AUC-gap results.Source: Figure 9 · p. 25, §3.5 · pp. 25–26

3.6 Non-inferiority analysis of the mitigation strategies

  • Five mitigation configurations satisfied the joint criterion on both datasets; QSVM with GRL+EO reduced disparity most among admissible DDI configurations and quantum CheXpert configurations while improving AUC in both.The criterion required reduced disparity and no more than a 0.02-point loss in global AUC. QCNN with GRL on DDI instead had the experiment’s largest disparity reduction (δ∆ = −0.148; absolute gap 0.011) but lost 0.042 AUC points; QCNN with GRL on CheXpert increased disparity despite strong aggregate performance. The criterion assesses global discrimination only, not deterioration within each subgroup, and CheXpert downstream AUC gaps remained 0.183–0.367.Source: §3.6 · pp. 26–27, Table 3 · p. 27
  • In CheXpert, none of the evaluated mitigation strategies brought subgroup performance to a level supporting responsible deployment, regardless of how disparity changed.The youngest age group had baseline AUCs close to chance for several downstream models.Source: §3.6 · p. 28
  • Non-inferiority is relative to a model’s baseline and does not certify that residual disparity is clinically acceptable.A configuration may satisfy the criterion while degrading one subgroup if the global AUC loss is offset elsewhere.Source: §3.6 · p. 28
  • The study did not compute subgroup-level AUCs for all mitigated settings, so it could not establish whether the worst-performing group improved or the best-performing group deteriorated.Distinguishing these mechanisms was identified as the most important analytical extension for future work.Source: §3.6 · p. 28

4 Limitations and Future Work

The audit is simulation-based and uses fixed ImageNet-pretrained ResNet-18 embeddings, limiting hardware and quantum-specific attribution. Its clinical interpretation is also constrained by task scope, uncertainty estimates, and the absence of subgroup-level results for mitigated models.

4 Limitations and Future Work

  • Quantum models ran only in PennyLane classical simulation, so these results do not establish real-hardware clinical feasibility.Device noise, connectivity, finite sampling, calibration drift, and execution conditions are not captured.Source: §4 · p. 28
  • ResNet-18 ImageNet embeddings were compressed to ten principal components before quantum processing, and no matched-capacity classical control on that representation was performed.Observed fairness behaviour therefore cannot be cleanly attributed to the quantum component; the study supports conclusions about hybrid classical–quantum pipelines, not quantum advantage.Source: §4 · pp. 28–29
  • The quantum models were restricted to ten-qubit simulations and compact projected features, limiting their dimensionality and expressiveness.The constraint kept QSVM kernel evaluation and QCNN training computationally feasible.Source: §4 · p. 29
  • The audit covers two binary imaging tasks and two sensitive attributes, not the wider range of demographic, acquisition, institutional, and clinical factors relevant to medical AI fairness.The authors call for larger and more diverse datasets, intersectional groups, additional attributes, and external validation cohorts.Source: §4 · p. 29
  • Uncertainty estimates reflect five random seeds, not metric-level resampling confidence intervals, and configurations were not compared with paired tests.This especially limits subgroup inference and close comparisons such as QSVM with GRL+EO versus PanDerm on DDI.Source: §4 · p. 29
  • The non-inferiority criterion constrains global AUC only; subgroup-level AUC changes for mitigated settings were not computed.Applying the margin to each subgroup would distinguish improvement in the worst-performing group from deterioration in the best-performing group.Source: §4 · pp. 29–30
  • The reported metrics do not characterize clinical consequences of errors or weight false negatives and false positives differently.The authors recommend subgroup sensitivity, specificity, and predictive values at clinically motivated operating points, plus net-benefit analysis.Source: §4 · p. 30
  • GRL treats sensitive-attribute information as undesirable without separating demographic shortcuts from legitimate epidemiological signal, and the evaluated mitigations did not uniformly improve fairness metrics.Because pleural-effusion prevalence rises with age in CheXpert, age-invariance objectives may remove clinically informative structure; the tested methods were differentiable EO, GRL, and their combination.Source: §4 · p. 30

5 Conclusions

Across DDI and CheXpert-derived data, the audit finds that QML models are not inherently fair and that mitigation effects vary by model and dataset. Reduced disparity must be weighed against preserved diagnostic performance, while the quantum models’ classical-simulation results do not establish quantum advantage.

5 Conclusions

  • Quantum classifiers were not inherently fair: QSVM with GRL+EO was the strongest non-foundation performer on DDI, while QCNN with GRL led non-foundation performance on CheXpert, where EVA-X remained best.QSVM with GRL+EO approached PanDerm in the DDI performance–fairness trade-off; QCNN with GRL increased disparity on CheXpert.Source: §5 · p. 31
  • EO, GRL, and their combination reduced disparity in some settings but preserved or increased it in others, and AUC-gap and demographic-parity comparisons could yield different interpretations.The metric differences were especially relevant when class prevalence varied across sensitive groups.Source: §5 · p. 31
  • On DDI, QSVM with GRL+EO met the non-inferiority criterion on both datasets, while QCNN with GRL had the experiment’s smallest disparity but lost 0.042 AUC points against its DDI baseline.QSVM with GRL+EO improved global discrimination on both datasets; on CheXpert, SVM with GRL achieved a larger admissible disparity reduction.Source: §5 · p. 31
  • The audit concludes that fairness-gap reductions are not intrinsically beneficial; clinical benefit requires reducing unjustified disparity without meaningful diagnostic loss or increased patient harm.The audit lacks per-subgroup changes against baseline, metrics at clinically relevant operating points, and net-benefit analysis; demographic invariance may also remove legitimate age-related epidemiological information.Source: §5 · pp. 31–32
  • The work reports being among the first empirical studies to apply GRL and combined GRL+EO mitigation in QML-based medical image classification.It recommends evaluating quantum classifiers under the subgroup-aware responsible-AI protocols expected of classical medical AI.Source: §5 · p. 32
  • The quantum models ran in classical simulation on ten principal components from a classical ImageNet-pretrained backbone, so the study does not demonstrate quantum advantage.Without a matched-capacity classical control using the same representation, the quantum component cannot be isolated from feature extraction, dimensionality reduction, capacity, or regularisation differences.Source: §5 · p. 32

Related papers

Loading 2610.37819v1…