Matched but Not Better: An Equivalence-Tested Benchmark of Quantum Machine Learning on Six Clinical Datasets and the QML-Health-READY Reporting Framework
Organizations: School of Engineering, St Aloysius (Deemed to be University), Mangalore 575003, Karnataka, India · School of ICT, Bahrain Polytechnic, Manama, Bahrain.
Abstract
Background: Quantum machine learning (QML) is promoted for clinical prediction, but comparisons with classical models are rarely paired or tested for equivalence. Methods: We compared a quantum-kernel support vector machine (QSVM; 4 qubits, tuned bandwidth) and a variational quantum classifier (VQC) with five tuned classical models across six public clinical datasets. All models used identical four-component inputs and 25 cross-validation splits. Paired AUROC differences were assessed using a corrected resampled t-test, Holm-adjusted equivalence tests (margins: +/-0.02 and +/-0.05), and random-effects pooling. Ablation, noise, finite-shot, and implementation-fault analyses were conducted. We propose QML-Health-READY, an eight-domain reporting framework. Results: Neither quantum model significantly outperformed any classical comparator. Pooled across datasets, QSVM differed from the kernel-matched radial basis function (RBF) SVM by -0.006 AUROC (95% CI: -0.017 to +0.004), equivalent to +/-0.02. VQC differed from the best classical model by -0.007 (95% CI: -0.019 to +0.004). QSVM was equivalent to RBF SVM within +/-0.05 in four of six datasets but was significantly worse than the best classical model on NHANES 2017-18 (-0.076). A classically trivial product-state kernel performed as well as the entangling kernel. With depolarizing noise (p = 0.01) and 1,024 shots per kernel entry, QSVM AUROC decreased by 0.054 (95% CI: 0.036 to 0.071). Untuned bandwidth produced near-chance AUROC (0.48-0.59). Conclusions: Correctly configured quantum classifiers approached but did not exceed tuned classical models. Finite-shot estimation, rather than gate noise, was their main practical cost, while configuration errors could reverse conclusions. QML-Health-READY defines reporting requirements for credible claims.
Figures & tables
| Domain | Name | Items |
|---|---|---|
| R† | Rationale | Hypothesised mechanism for advantage; data-complexity analysis (geometric difference, kernel alignment) [31,40]. |
| E† | Encoding | Encoding, feature scaling, kernel bandwidth and its selection procedure (checking grid edges), dimensionality reduction fitted within training folds, qubit count [21,23]. |
| A† | Algorithm & baselines | ≥ 4 tuned classical baselines including a kernel-matched classical model, given the same inputs and tuning budget; entanglement-ablated quantum model [11]. |
| D† | Data & splits | Provenance, sample size, class balance, subject-level splitting, leakage audit (including post-baseline predictors). |
| Y† | Yield | Repeated nested cross-validation with identical splits for all models; paired tests corrected for split dependence [41,42]; equivalence testing with a pre-specified margin when parity is claimed [43,44]; multiplicity control. |
| R† | Robustness | Explicit noise model (verified to be active) and finite-shot estimation; hardware-calibrated noise or real hardware where feasible [20,36]. |
| Dataset | Domain | Source | n | Predictors | Class 1 | Prevalence |
|---|---|---|---|---|---|---|
| Breast Cancer Wisconsin | Oncology | scikit-learn / UCI | 569 | 30 | benign tumour | 63% |
| Heart Disease Cleveland | Cardiology | UCI | 297 | 13 | any heart disease | 46% |
| Pima Indians Diabetes | Metabolic | UCI (mirror) | 768 | 8 | diabetes | 35% |
| Heart Failure | Cardiology | UCI | 299 | 11 | death during follow-up | 32% |
| TCGA-BRCA | Genomic oncology | UCSC Xena | 780 | 200 | ER-positive | 77% |
| NHANES 2017–18 | Epidemiology | CDC NHANES | 2,000 | 6 | diabetes | 18% |
| Model | Breast Cancer Wisconsin | Heart Disease Cleveland | Pima Indians Diabetes | Heart Failure | TCGA-BRCA | NHANES 2017–18 |
|---|---|---|---|---|---|---|
| Logistic regression | 0.993 0.002 | 0.898 0.022 | 0.815 0.013 | 0.710 0.029 | 0.954 0.009 | 0.777 0.012 |
| SVM (RBF) | 0.994 0.002 | 0.892 0.022 | 0.820 0.013 | 0.736 0.028 | 0.952 0.010 | 0.731 0.015 |
| SVM (polynomial) | 0.993 0.004 | 0.889 0.022 | 0.817 0.014 | 0.712 0.019 | 0.958 0.008 | 0.712 0.020 |
| Random forest | 0.985 0.004 | 0.880 0.023 | 0.812 0.014 | 0.709 0.027 | 0.946 0.011 | 0.773 0.012 |
| Gradient boosting | 0.985 0.004 | 0.870 0.022 | 0.812 0.014 | 0.664 0.025 | 0.944 0.008 | 0.770 0.012 |
| QSVM | 0.990 0.003 | 0.884 0.024 | 0.807 0.014 | 0.708 0.028 | 0.956 0.010 | 0.700 0.016 |
| Dataset | Comparison | AUROC [95% CI] | p(t) | p(TOST 0.02) | p(TOST 0.05) | Verdict 0.02 / 0.05 |
|---|---|---|---|---|---|---|
| Breast Cancer Wisconsin | QSVM-SVM (RBF) | -0.004 [-0.011, +0.003] | 0.923 | <0.001 | <0.001 | E / E |
| Breast Cancer Wisconsin | QSVM-best (SVM (RBF)) | -0.004 [-0.011, +0.003] | 0.569 | <0.001 | <0.001 | E / E |
| Breast Cancer Wisconsin | VQC (tuned)-best (SVM (RBF)) | -0.003 [-0.006, +0.001] | 0.473 | <0.001 | <0.001 | E / E |
| Heart Disease Cleveland | QSVM-SVM (RBF) | -0.007 [-0.034, +0.019] | 1.000 | 0.658 | 0.004 | I / E |
| Heart Disease Cleveland | QSVM - best (Logistic regression) | -0.013 [-0.034, +0.007] | 0.569 | 0.750 | 0.002 | I / E |
| Heart Disease Cleveland | VQC (tuned) - best (Logistic regression) | -0.026 [-0.047, -0.004] | 0.136 | 1.000 | 0.029 | I / E |
| Domain | Item tested | Measured effect in this study |
|---|---|---|
| R† Rationale | Data-complexity analysis | Geometric difference 2.4–4.2 against a scale of ≈17, and lower quantum than RBF kernel–target alignment on all six datasets: no data-driven advantage was expected, and none was seen. |
| E† Encoding | Bandwidth tuning and input scaling | Fixing the bandwidth at 1 lowered QSVM AUROC by 0.18–0.45, to near chance. |
| A† Algorithm & baselines | Kernel-matched comparator; entanglement ablation | The NHANES deficit was -0.076 against the best classical model but -0.031 (not significant) against the RBF SVM; the product-state kernel matched the entangling one. |
| D† Data & splits | Leakage audit (post-baseline predictors) | Including heart-failure follow-up time inflated classical AUROC by 0.09–0.12. |
| Y† Yield | Pre-specified equivalence margin | QSVM–RBF SVM equivalence held on 4 of 6 datasets at 0.05 but 2 at 0.02 and 0 at 0.01. |
| R† Robustness | Explicit noise and finite shots | Exact depolarising noise (p = 0.01) changed AUROC by at most 0.007; 1,024 shots lowered it by a pooled 0.054. |
Quantum Machine Learning
Quantum machine learning uses quantum circuits and quantum data to build learning algorithms and study their computational and statistical limits.
See all 267 papers