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

cs.LGOct 27, 2026

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

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

Organizations: Faculty of Sciences, S˜ao Paulo State University (Unesp), Bauru, Brazil. · Laboratory of Magnetic Resonance in Neuroradiology (LIM-44), University of S˜ao Paulo Faculty of Medicine Clinics Hospital, S˜ao Paulo, Brazil. · Hospital Israelita Albert Einstein, S˜ao Paulo, Brazil.

Abstract

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.

Figures & tables

Figure 1: Bias auditing pipeline for classical and quantum machine learning in medical imaging. The workflow starts from two medical imaging datasets: DDI for dermatology and CheXpert for chest radiography. Sensitive attributes are defined using Fitzpatrick skin type groups for DDI and age groups for CheXpert. Images are encoded using a ResNet-18 model pretrained on ImageNet, and the resulting embeddings are used as inputs to classical and quantum classifiers. The models are evaluated under baseline, EO, GRL, and GRL+EO settings, followed by subgroup-aware analysis using predictive and fairness metrics.

Figure 2: Class distribution across sensitive groups in the evaluated datasets. Panel (a) shows the percentage distribution of benign and malignant cases across Fitzpatrick skin type groups in DDI. Panel (b) shows the percentage distribution of no pleural effusion and pleural effusion cases across age groups in CheXpert. The figure highlights subgroup-specific differences in class prevalence, supporting the need for fairness-aware evaluation across sensitive groups.

Figure 3: Fairness mitigation strategies used in the proposed evaluation. Panel (a) shows Gradient Reversal Learning (GRL), in which a shared encoder produces a representation used by both the clinical classifier and a sensitive-attribute adversary. The gradient reversal layer gradually reverses the adversarial gradient according to a scheduled coefficient, encouraging the representation to preserve disease-relevant information while reducing information about the sensitive group. Panel (b) shows the differentiable Equalized Odds (EO) penalty used during training. The predicted probabilities are separated by true class and sensitive group, group-wise mean probabilities are computed within each class, and the variance of these group-wise means is added to the binary cross-entropy loss. This encourages the model to reduce group-dependent differences in prediction behaviour for both negative and positive cases.

Model	Mitigation	AUC	Balanced acc.	F1-score
EVA-X (foundation)	—	$0.925\pm 0.029$	$0.846\pm 0.032$	$0.782\pm 0.043$
SVM	Baseline	$0.783\pm 0.032$	$0.732\pm 0.031$	$0.611\pm 0.044$
SVM	Eq. Odds	$0.797\pm 0.030$	$0.735\pm 0.031$	$0.617\pm 0.045$
SVM	GRL	$0.783\pm 0.031$	$0.737\pm 0.030$	$0.618\pm 0.044$
SVM	GRL+EO	$0.789\pm 0.032$	$0.709\pm 0.031$	$0.582\pm 0.044$
MLP	Baseline	$0.805\pm 0.029$	$0.753\pm 0.031$	$0.640\pm 0.046$

Table 1: Global predictive performance on CheXpert across models and fairness mitigation strategies. Values are reported as point estimates with standard deviations across seeds when available. The foundation model is reported as a reference baseline and marked with “—” because no fairness mitigation strategy was applied. Bold values indicate the highest point estimate among non-foundation models for each metric. Bold marks the highest point estimate only; because the reported deviations are seed-to-seed variability rather than confidence intervals, and are not paired across models, differences between configurations with overlapping deviations should not be read as evidence of superiority Altman and Bland (1995) .

Model	Mitigation	AUC	Balanced acc.	F1-score
PanDerm (foundation)	—	$0.814\pm 0.051$	$0.766\pm 0.043$	$0.640\pm 0.064$
SVM	Baseline	$0.771\pm 0.049$	$0.724\pm 0.042$	$0.568\pm 0.061$
SVM	Eq. Odds	$0.761\pm 0.050$	$0.704\pm 0.047$	$0.549\pm 0.065$
SVM	GRL	$0.755\pm 0.048$	$0.687\pm 0.045$	$0.533\pm 0.068$
SVM	GRL+EO	$0.733\pm 0.054$	$0.654\pm 0.045$	$0.484\pm 0.078$
MLP	Baseline	$0.776\pm 0.047$	$0.693\pm 0.046$	$0.543\pm 0.072$

Table 2: Global predictive performance on DDI across models and fairness mitigation strategies. Values are reported as point estimates with standard deviations across seeds when available. The foundation model is reported as a reference baseline and marked with “—” because no fairness mitigation strategy was applied. Bold values indicate the highest point estimate among non-foundation models for each metric. As in Table 1 , overlapping deviations do not support claims of superiority between configurations.

Figure 4: Baseline ROC curves on DDI by Fitzpatrick skin type group. Each panel shows the overall ROC curve and subgroup-specific curves for FST I–II, FST III–IV, and FST V–VI. The figure illustrates that global predictive performance may conceal subgroup-level variation in model discrimination.

Figure 5: Baseline ROC curves on CheXpert by age group. Each panel shows the overall ROC curve and subgroup-specific curves for the 0–40, 41–60, and 60+ age groups. The downstream classifiers show weaker discrimination for the youngest subgroup, highlighting the importance of subgroup-aware evaluation.

Figure 6: Performance–fairness trade-off across models and mitigation strategies. Each point represents a model and mitigation setting. The horizontal axis reports mean AUC, while the vertical axis reports AUC gap across sensitive groups. Colour indicates the mitigation strategy, marker shape indicates model type, and bubble size represents balanced accuracy. Downward movement alone does not imply clinical benefit; see Section 3.6 for the joint criterion.

Figure 7: AUC gap summary across models and mitigation strategies. The heatmap reports the AUC gap between the best-performing and worst-performing sensitive groups. Lower values indicate smaller subgroup disparity. The two panels compare DDI and CheXpert under baseline and fairness mitigation settings. The heatmap reports disparity only; it should be read together with Table 3 , which reports the accompanying change in global discrimination.

Figure 8: Performance–fairness trade-off using demographic parity gap. Each point represents a model and mitigation setting. The horizontal axis reports mean AUC, while the vertical axis reports demographic parity gap across sensitive groups. Colour indicates the mitigation strategy, marker shape indicates model type, and bubble size represents balanced accuracy.

Figure 9: Demographic parity gap summary across models and mitigation strategies. The heatmap reports the demographic parity gap across sensitive groups. Lower values indicate smaller differences in positive prediction rates. The two panels compare DDI and CheXpert under baseline and fairness mitigation settings.

Model	Mitigation	$\bm{\delta_{\mathrm{AUC}}}$	$\bm{\delta_{\Delta}}$	Acceptable
DDI (sensitive attribute: Fitzpatrick skin type)
SVM	Eq. Odds	$-0.010$	$-0.058$	Yes
SVM	GRL	$-0.016$	$-0.076$	Yes
SVM	GRL+EO	$-0.038$	$+0.057$	No
MLP	Eq. Odds	$-0.007$	$-0.030$	Yes
MLP	GRL	$+0.005$	$-0.067$	Yes

Table 3: Joint assessment of disparity reduction and non-inferiority for each mitigated configuration, relative to the unmitigated baseline of the same model family.

\delta_{\mathrm{AUC}}

is the change in global AUC (Eq. 18 ) and

\delta_{\Delta}

is the change in subgroup AUC gap (Eq. 19 ); negative

\delta_{\Delta}

indicates reduced disparity. A configuration is marked “Yes” only when it reduces disparity (

\delta_{\Delta}<0

) and does not lose more than

\varepsilon=0.02

AUC points (

\delta_{\mathrm{AUC}}\geq-0.02

). Values are derived from Tables 1 and 2 and from Figure 7 . Bold marks the largest admissible disparity reduction in each dataset.

Fairness in Medical Image Classification

See all 21 papers

Comparative Study of Quantum and Classical Machine Learning Models in Binary Classification
Sep 20, 2026Anand Kumar Mishra, Ramanuj AwasthiQuantum Machine Learning Quantum Kernel Methods
School of Engineering & Technology (UIET), CSJM University, Kanpur, Uttar Pradesh, India
Quantum Kernel Advantage over Classical Collapse in Medical Foundation Model Embeddings
Apr 27, 2026Sebastian Cajas Ordóñez, Felipe Ocampo Osorio, Dax Enshan Koh +10Chest X-Ray Classification Healthcare
1MIT Critical Data, Massachusetts Institute of Technology, Cambridge, MA, USA · 2Clinical Research Center, Artificial Intelligence Unit, Fundaci´on Valle del Lili, Cali, Valle del Cauca, Colombia · 3Quantum Innovation Centre (Q.InC), Agency for Science, Technology and Research (A*STAR), 2 Fusionopolis Way, Innovis #08-03, Singapore 138634, Republic of Singapore +15
Quantum Diffusion Models for Medical Image Analysis
Sep 25, 2026Francesco Aldo Venturelli, Stefano Martina, Marco Parigi +3Quantum Generative Modeling Diffusion Models
BCN Medtech, Department of Engineering, Universitat Pompeu Fabra, Barcelona, Spain · Barcelona Supercomputing Center, Barcelona, Spain · Department of Physics and Astronomy, University of Florence, Sesto Fiorentino, Florence, Italy +2

Sep 20, 2026J/K move · Enter open · S save

Comparative Study of Quantum and Classical Machine Learning Models in Binary Classification

Anand Kumar Mishra, Ramanuj Awasthi

School of Engineering & Technology (UIET), CSJM University, Kanpur, Uttar Pradesh, India

Quantum Machine Learning Quantum Kernel Methods Binary Classification Variational Quantum Circuits

Open paper AskPDF arXiv

A potential path forward is Quantum Machine Learning (QML), which aims to leverage quantum computing in conjunction with classical machine learning to enhance computing efficiency and the expressiveness of models. In this paper, two different quantum classifiers - Variational Quantum Classifier (VQC) and Quantum Kernel Support Vector Machine (QSVM) - are compared with three classical classifiers as baseline classifiers - Logistic Regression, Support Vector Machine (SVM), and a Multi-Layer Perceptron (MLP) - on the Breast Cancer Wisconsin dataset. The quantum circuits were created in the PennyLane framework and simulated on a classical backend. However, in terms of accuracy, classical Logistic Regression performed better with an accuracy of 97.8%, classical SVM and QSVM with an accuracy of 95.6% each, although the Quantum VQC achieved a lower accuracy of 88.9% and had a recall of 100% for the benign class, though it correctly identified only 12 of the 17 malignant cases (a malignant-class recall of approximately 70.6%). The drawback of quantum models is the higher training time; however, since the quantum circuit needs to be classically simulated, the quantum SVM took 23.29 seconds compared to less than 0.01 seconds for the classical linear models. These results indicate that for small structured datasets, classifiers based on quantum computing have not yet surpassed well-tuned classical counterparts. In some respects (e.g., benign-class recall), they perform competitively, though not on malignant-class recall, where the VQC in particular performed worse than the classical baselines, which is worth further investigation on real quantum computers.

Abstract

Fairness in Medical Image Classification

Explore similar work

Comparative Study of Quantum and Classical Machine Learning Models in Binary Classification

Quantum Kernel Advantage over Classical Collapse in Medical Foundation Model Embeddings

Quantum Diffusion Models for Medical Image Analysis