Quantum kernel methods are increasingly applied to clinical prediction, but reported advantages rarely come with ablations, confidence intervals or negative controls. We benchmark a hybrid quantum-classical stacking model under nested cross-validation on three public biomedical cohorts: CRyPTIC tuberculosis drug resistance (12,287 isolates, six drugs), Duke DCE-MRI radiogenomics (n = 922, five endpoints) and MIAS mammography (322 images). Against an all-classical ensemble, the quantum branches change AUC by between -0.0024 and +0.0048 on every tuberculosis and Duke endpoint. Of twenty-four paired comparisons none significantly favours the hybrid, and the one nominally significant difference (p = 0.012, the hybrid below logistic regression on mammography) does not survive correction for multiplicity. Two measurements explain the null: fidelity-kernel off-diagonals decay as 0.5037^n (R^2 = 0.99), and a 277-fold increase in simulated entanglement moves AUC by 0.0148. We recast the inductive-bias argument for quantum kernels as a screen that needs no quantum computation: a quantum feature map can help only if the target depends on high-order feature interactions. The screen measured all six tuberculosis drugs as first-order before any quantum model was run. On synthetic controls an entangling map overtakes tuned classical baselines only at maximal interaction order, and its margin grows with sample size, against the common small-data claim. We also document three evaluation failures in our own pipeline, including a feature-label leak (0.918 AUC where the rebuild gives 0.54-0.59) and stacking leakage worth up to 0.09 AUC. All features, predictions and code are released. Data: https://doi.org/10.5281/zenodo.23344459 . Code: https://doi.org/10.5281/zenodo.23344193 .
Quantum kernel methods are increasingly applied to clinical prediction, but reported advantages rarely come with ablations, confidence intervals or negative controls. We benchmark a hybrid quantum-classical stacking model under nested cross-validation on three public biomedical cohorts: CRyPTIC tuberculosis drug resistance (12,287 isolates, six drugs), Duke DCE-MRI radiogenomics (n = 922, five endpoints) and MIAS mammography (322 images). Against an all-classical ensemble, the quantum branches change AUC by between -0.0024 and +0.0048 on every tuberculosis and Duke endpoint. Of twenty-four paired comparisons none significantly favours the hybrid, and the one nominally significant difference (p = 0.012, the hybrid below logistic regression on mammography) does not survive correction for multiplicity. Two measurements explain the null: fidelity-kernel off-diagonals decay as 0.5037^n (R^2 = 0.99), and a 277-fold increase in simulated entanglement moves AUC by 0.0148. We recast the inductive-bias argument for quantum kernels as a screen that needs no quantum computation: a quantum feature map can help only if the target depends on high-order feature interactions. The screen measured all six tuberculosis drugs as first-order before any quantum model was run. On synthetic controls an entangling map overtakes tuned classical baselines only at maximal interaction order, and its margin grows with sample size, against the common small-data claim. We also document three evaluation failures in our own pipeline, including a feature-label leak (0.918 AUC where the rebuild gives 0.54-0.59) and stacking leakage worth up to 0.09 AUC. All features, predictions and code are released. Data: https://doi.org/10.5281/zenodo.23344459 . Code: https://doi.org/10.5281/zenodo.23344193 .
Quantum kernel methods are the most concrete near-term proposal for quantum machine learning on classical data, yet published benchmarks disagree on whether they outperform tuned classical kernels, and the disagreement cannot be resolved from the literature because the comparisons are not protocol-controlled. We argue that the obstacle is one of measurement, not theory: the usual setup fixes the quantum feature map's data scaling while grid-searching the classical kernel's hyperparameters, recomputes nothing per fold, and reports cost in simulator time. We therefore treat the comparison itself as the object of study and impose three controls. The bandwidth is tuned on the same grid budget as the classical γ, the kernel is recomputed inside every fold from fold-fitted preprocessing, and cost is counted in hardware circuit evaluations. On top of these we run a factorial design of 324 cells in which dimensionality, training size, class separability and data structure are explicit covariates, together with four standard datasets, a shots-versus-gate-noise ablation, and a cost sweep. The dominant effect we measure is not quantum: tuning the feature map's bandwidth is worth 0.120 accuracy on synthetic data and 0.339 on real data, exceeding every quantum-classical difference we observe. Tuned quantum kernels are then statistically indistinguishable from a tuned RBF on all four real datasets. The advantage on synthetic data is confined almost entirely to data generated by the feature map itself, and it shrinks with dimensionality (-0.0045 per qubit) while growing with sample size. Finite sampling costs an order of magnitude more accuracy than gate and readout noise, and a fidelity kernel needs 5×107 circuits at n = 10,000. Protocol variance is thus sufficient to explain the field's contradictory conclusions.
Background: Breast cancer remains a leading cause of cancer-related mortality, and reliable computational decision support is increasingly viewed as a complement to expert pathological assessment rather than a replacement for it. Variational quantum classifiers (VQCs) and Quantum Support Vector Machines (QSVMs) have recently been promoted as candidate models for medical classification, yet most published comparisons rely on internal hold-out validation alone and report only a single point estimate of discrimination, omitting calibration, decision-analytic value, and explainability—three ingredients that any clinically credible model must furnish. Methods: We assembled a complete quantum–classical machine learning pipeline and evaluated it under a deliberately stringent protocol designed to expose, rather than conceal, the limitations of current Noisy Intermediate-Scale Quantum (NISQ)-era models. The analytical hypothesis was conservative and stated in advance; in light of saturated classical baselines on this benchmark, we did not anticipate a quantum advantage in raw discrimination, and we framed the study as a methodological probe rather than as a competition. Using the Wisconsin Diagnostic Breast Cancer (WDBC) dataset (n = 569) for development and an independent Wisconsin Original (WBC) cohort (n = 683) for external validation, we benchmarked five classical learners (XGBoost, LightGBM, CatBoost, RandomForest, RBF-SVM), two quantum models (an eight-qubit VQC implemented in PennyLane and a ZZ-feature-map QSVM implemented in Qiskit), and a stacked hybrid ensemble. The evaluation framework combined Optuna-driven hyperparameter optimisation, internal–external cross-validation, and external validation on the independent WBC cohort. Robustness and interpretability were then probed through circuit depth and embedding rotation ablation, depolarising noise stress tests, learning curve and feature stability analysis, decision curve analysis, and dual SHAP-based explanations covering both a direct tree-based explanation and a quantum surrogate. Reporting followed the TRIPOD + AI guideline. Results: On the internal test partition, RBF-SVM achieved the highest discrimination (AUC = 0.998), with XGBoost, LightGBM, CatBoost, the hybrid ensemble, and the VQC clustering between 0.992 and 0.996; the QSVM with a ZZ-fidelity kernel underperformed substantially (AUC = 0.727). Pairwise tests for correlated ROC curves indicated that most differences among top models were not statistically significant. On the external WBC cohort, model rankings reorganised, as RBF-SVM (AUC = 0.986, 95% CI 0.946–0.997), RandomForest (0.985, 95% CI 0.945–0.996), VQC (0.983, 95% CI 0.942–0.995), and the hybrid ensemble (0.982, 95% CI 0.941–0.995) all retained near-ceiling discrimination with extensively overlapping confidence intervals. Ablation analysis demonstrated that the choice of embedding rotation is decisive—Z-rotation embeddings collapsed VQC performance to chance levels (AUC ≈ 0.50), whereas X- and Y-rotations preserved it. Depolarising noise up to p = 0.10 had a negligible effect on the VQC, and SHAP analyses converged on worst concave points, mean concave points, and worst area as the dominant predictors across both classical and quantum models. Decision curve analysis showed positive net benefit for both classical and hybrid models across the clinically meaningful threshold range, exceeding both the treat-all and treat-none reference strategies throughout. Conclusions: In the present regime, the principal contribution of QML is not raw discrimination—modern classical learners are already at the data ceiling—but the construction of a rigorous, reproducible, externally validated, and interpretable benchmarking framework in which quantum models can be fairly compared with their classical counterparts. Because evaluation was confined to curated benchmark datasets rather than real-world clinical populations, the interpretability and net benefit findings reported here should be read as benchmark-level evidence and not as a demonstration of readiness for clinical deployment.
Do quantum kernels improve cross-sectional stock return prediction? We run a controlled horse race on the Chinese A-share market in which a quantum fidelity kernel, a projected quantum kernel, and a classical RBF control share identical training subsamples, solver, and tuning budgets, so that only the kernel is exchanged. On the main evaluation -- a point-in-time universe and 170 walk-forward windows (2012-2025) -- no quantum advantage exists: the fidelity kernel is indistinguishable from its RBF control (ΔIC =+0.005, p=0.42), and a 2×2 design crossing kernel type with training budget (a Nystrom extension to the full ~38,000-observation windows) shows quantum kernels matching, but never beating, equal-budget linear models; after family-wise correction no pairwise difference among eleven models is significant, with point estimates favoring penalized linear regressions throughout. We then document how the opposite conclusion arises: a 60-window evaluation on a universe screened with full-sample information makes the same quantum kernel appear dominant on stability criteria and significantly better than neural baselines. Interaction characteristics from the anomalies literature help nothing, quantum or classical; a widened bandwidth grid reveals an interior optimum rather than the near-classical endpoint a coarse grid suggests; and the geometric difference, while large throughout (g≫1), does not predict out-of-sample gains (ρ=−0.20). We propose protocol standards -- kernel-swap controls, budget-equalized comparisons, point-in-time universes, and multiplicity-robust inference -- for empirical claims of quantum advantage in finance.
Resting-state functional connectivity is widely used in machine learning-based classification of neuropsychiatric disorders. Quantum machine learning models have been proposed for functional connectivity data, but few studies have evaluated them under leave-site-out validation or against matched classical models. We compared six quantum model families with classical ones using identical correlation-derived features across four public cohorts (4,308 subjects). Leave-site-out validation was used for the Autism Brain Imaging Data Exchange (ABIDE I), ADHD-200 (attention-deficit/hyperactivity disorder), and REST-meta-MDD (major depressive disorder) cohorts, while the UCLA Consortium for Neuropsychiatric Phenomics (UCLA-CNP) was evaluated using stratified 10-fold cross-validation. The quantum kernel was statistically indistinguishable from a matched classical kernel across all four cohorts, with the average accuracy difference within 1.4 percentage points at 95% confidence. On ABIDE I, a hybrid quantum convolutional network achieved similar accuracy to its classical counterpart (0.645 vs. 0.641), while a trainable quantum kernel exceeded its classical counterpart by 0.064 in area under the curve (AUC). Quantum convolutional networks also achieved higher AUC than their matched classical networks on this cohort and showed smaller train-test gaps on all four cohorts, without an accuracy gain. Overall, most quantum models matched their classical counterparts, but none consistently outperformed them. This study, among the first quantum evaluations of ABIDE I under leave-site-out validation, provides a feature-matched benchmark of quantum and classical models comprising 284 matched comparisons. We further analyse fidelity-based graph construction and account for the observed quantum-classical parity through effective-dimension and eigenspectrum analyses.
Quantum kernel methods have been proposed as a promising approach for leveraging near-term quantum computers for supervised learning, yet rigorous benchmarks against strong classical baselines remain scarce. We present a comprehensive empirical study of quantum kernel support vector machines (QSVMs) across nine binary classification datasets, four quantum feature maps, three classical kernels, and multiple noise models, totalling 970 experiments with strict nested cross-validation.
Our analysis spans four phases: (i) statistical significance testing, revealing that none of 29 pairwise quantum-classical comparisons reach significance at α=0.05; (ii) learning curve analysis over six training fractions, showing steeper quantum slopes on six of eight datasets that nonetheless fail to close the gap to the best classical baseline; (iii) hardware validation on IBM ibm_fez (Heron r2), demonstrating kernel fidelity r≥0.976 across six experiments; and (iv) seed sensitivity analysis confirming reproducibility (mean CV 1.4%).
A Kruskal-Wallis factorial analysis reveals that dataset choice dominates performance variance (ε2=0.73), while kernel type accounts for only 9%. Spectral analysis offers a mechanistic explanation: current quantum feature maps produce eigenspectra that are either too flat or too concentrated, missing the intermediate profile of the best classical kernel, the radial basis function (RBF). Quantum kernel training (QKT) via kernel-target alignment yields the single competitive result -- balanced accuracy 0.968 on breast cancer -- but with ~2,000x computational overhead.
Our findings provide actionable guidelines for quantum kernel research. The complete benchmark suite is publicly available to facilitate reproduction and extension.
Background/Objectives: Hybrid quantum–classical neural networks are reported to classify brain tumors on magnetic resonance imaging (MRI) accurately, but these reports rarely control for data leakage or test whether the quantum circuit contributes. We examined both issues. Methods: We audited version 2 of the Kaggle Brain Tumor MRI Dataset (7200 images) with mirror- and rotation-invariant duplicate detection and by matching every image to its source collections, which recovered patient identities for 4846 tumor images. A variational quantum circuit (VQC) was compared with size-matched, deeper and Fourier-surrogate classical controls using frozen features, end-to-end fine-tuning and training from scratch, all under patient-grouped five-fold cross-validation. Results: Of the 1600 test images, 1269 (79%) had a copy in the training partition, and 915 of the 917 figshare-derived test images shared a patient with it. On frozen features, a six-qubit VQC head was 6.0 percentage points (pp) less accurate than its classical control (95% confidence interval −6.8 to −5.2) and less accurate than a classical surrogate of its own function class. With fine-tuning, it stayed within about 1 pp of its control but was 0.7 pp less accurate (−1.2 to −0.1) after convergence-based training; a hybrid network trained from scratch was 8.3 pp less accurate than its control. Patient-grouped evaluation lowered accuracy by 3–8 pp relative to our earlier duplicate-grouped evaluation, and 76 glioma test images from another source remained poorly recognized. Conclusions: The quantum layer added cost but no accuracy, and leakage and provenance mattered more than the classifier head.