Drug-induced liver injury (DILI) remains difficult to anticipate computationally because many clinically important cases are idiosyncratic and are only weakly determined by chemical structure alone. We compared a support vector quantum-kernel classifier (SVQK) with matched classical baselines for binary DILI prediction on a benchmark FDA/Therapeutics Data Commons dataset of 475 compounds. All reported models were restricted to the same 16-feature principal-component representation and evaluated with stratified 5-fold cross-validation. Under this constrained setting, the SVQK achieved a mean AUC of 0.85 ± 0.03, compared with 0.78 ± 0.04 for the strongest classical baseline, Random Forest, with corresponding accuracies of 82.0 ± 2.5% and$75.0 \pm 3.0 \%$, respectively. We interpret this as a proof-of-concept result: the quantum-kernel model was competitive with, and slightly better than, the tested classical models under a qubit-feasible feature budget, rather than evidence of general quantum superiority. All primary quantitative results were obtained from ideal and noisy simulations; limited runs on IBM hardware were used only for minor feasibility checks because current device noise precluded stable comparative benchmarking. Interpretability was limited to exploratory post-hoc analysis. Because the study uses a single 475-compound dataset without external validation, and because idiosyncratic DILI is notoriously hard to model, the findings should be regarded as preliminary.
Drug-induced liver injury (DILI) remains difficult to anticipate computationally because many clinically important cases are idiosyncratic and are only weakly determined by chemical structure alone. We compared a support vector quantum-kernel classifier (SVQK) with matched classical baselines for binary DILI prediction on a benchmark FDA/Therapeutics Data Commons dataset of 475 compounds. All reported models were restricted to the same 16-feature principal-component representation and evaluated with stratified 5-fold cross-validation. Under this constrained setting, the SVQK achieved a mean AUC of 0.85 ± 0.03, compared with 0.78 ± 0.04 for the strongest classical baseline, Random Forest, with corresponding accuracies of 82.0 ± 2.5% and75.0±3.0%, respectively. We interpret this as a proof-of-concept result: the quantum-kernel model was competitive with, and slightly better than, the tested classical models under a qubit-feasible feature budget, rather than evidence of general quantum superiority. All primary quantitative results were obtained from ideal and noisy simulations; limited runs on IBM hardware were used only for minor feasibility checks because current device noise precluded stable comparative benchmarking. Interpretability was limited to exploratory post-hoc analysis. Because the study uses a single 475-compound dataset without external validation, and because idiosyncratic DILI is notoriously hard to model, the findings should be regarded as preliminary.
Multi-morbidity is the co-occurrence of conditions such as diabetes, cardiac arrhythmia, and hypertension in a single patient—poses a risk stratification problem that classical ensemble models address through high-capacity feature interactions. Quantum kernel methods offer a structurally different encoding: classical inputs are mapped to quantum states via parameterized circuits, and similarity is computed as state fidelity rather than through conventional dot-product kernels. Empirical validation of this approach on tabular clinical data remains limited. This work compares quantum kernel support vector machines (SVMs) using ZZ and Pauli feature maps against Random Forest, XGBoost, and polynomial-kernel SVM baselines on a synthesized multi-morbidity dataset under a few-shot learning protocol spanning training sizes n ∈ {50, 100, 200, 500}. The central finding is training-size invariance in quantum kernels: both the ZZ and Pauli implementations maintained approximately 74% accuracy across all four training sizes, whereas Random Forest improved from 81.8 ± 2.0% at n = 50 to 88.7 ± 1.0% at n = 500, with XGBoost exhibiting a similar trend. This invariance constitutes the primary empirical contribution of the study. The results suggest that quantum feature maps determine the embedding geometry at circuit construction time, yielding consistent kernel representations regardless of the number of labeled training samples. In scenarios where labeled clinical data is structurally limited, such stability may provide practical value, although it comes at a 7–15 percentage-point accuracy deficit relative to optimized ensemble methods. Quantum kernel training time increased to several hundred seconds at n = 500, while all classical baselines remained below one second. Under classical statevector simulation, this computational overhead represents the principal constraint on near-term deployment.
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.
This study presents a comparative analysis of classical machine learning (ML) and quantum machine learning (QML) algorithms for industrial equipment fault detection using the AI4I 2020 Predictive Maintenance dataset. A total of nine models were evaluated through 5-fold stratified cross-validation: three classical models - Logistic Regression, Random Forest, and XGBoost - alongside six quantum models, namely Quantum Support Vector Machine (QSVM), Quantum K-Nearest Neighbors (QKNN), Quantum Kernel-based hybrid classifiers (QK+RF, QK+GB), Pegasos Quantum Support Vector Classifier (PegasosQSVC), and Variational Quantum Classifier (VQC). All quantum experiments were conducted using the IBM Qiskit framework and AerSimulator backend. Under a controlled comparison using an identical 300-sample budget and the same 5-fold cross-validation protocol, the best classical model (Random Forest, F1=0.902) outperformed the best quantum kernel method (QKNN, F1=0.820) by approximately 8.2 points. These results indicate that quantum kernel methods do not provide a dataefficiency advantage over classical methods on this dataset, although they achieve reasonably competitive results (F1 ≥0.75 for kernel-based methods) at substantially higher computational cost (quantum: ~124-137 s/fold for kernel methods, up to ~257 s/fold for VQC, vs. classical: 0.01-0.26 s/fold). We analyze the sources of this gap and attribute it primarily to a mismatch between the unitarily-invariant geometry of the fidelity kernel and the axisaligned threshold structure of the AI4I failure modes, rather than to information loss in dimensionality reduction, which the faircomparison design rules out. We further show that this apparent quantum competitiveness stems partly from an unequal-data-volume comparison common in prior QML benchmarking studies, which our fair-comparison protocol corrects.
Reproducibility archive (code + manuscript) for the paper "Benchmarking Quantum Kernel Support Vector Machines Against Classical Baselines on Tabular Data: A Rigorous Empirical Study with Hardware Validation" (arXiv:2604.18837). The archive contains the full Python benchmark framework, experiment configurations, unit tests, and the compiled manuscript. 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; (ii) learning curve analysis; (iii) hardware validation on IBM ibm_fez (Heron r2); and (iv) seed sensitivity analysis. 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 (RBF). The complete benchmark suite is publicly available to facilitate reproduction and extension.
It is an established practice in Ayurveda to classify human Prakriti from questionnaires, yet the machine learning aspect of it remains unexplored due to the availability of a dataset spanning over 40 to 132 categorical variables, with tension between dimensionality and small cohort size. This study directly compares Principal Component Analysis with derived quantum angle encoding to introduce the linear feature expansion by a quantum kernel to get a measurable classification advantage over PCA-based classical baselines on the same Prakriti dataset. The 105-feature dataset is preprocessed, ordinally encoded, standardized, and SMOTE-balanced. Eight PCA components are retained, capturing 37.9% of total variance; these form the shared representation for all models. Classical baselines (linear SVM, RBF SVM, Random Forest) operate directly on the eight PCA scores. The quantum pathway maps those scores to rotation angles and applies an analytic ZZ-kernel SVM. The ZZ-FeatureMap kernel is computed as a classical simulation, i.e., an analytic closed-form approximation evaluated in software, rather than executed on physical quantum hardware. A hybrid ensemble blends the quantum and classical predictions. The PCA SVM achieves 82.14% on the held-out test set, while the quantum kernel SVM reaches 98.21%. The hybrid ensemble achieves 98.21% with a five-fold cross-validated mean of 97.49%. The quantum feature map recovers the discriminative structure in the 256-dimensional ZZ-entanglement space that is absent from the 8-dimensional PCA subspace alone, with a 13.1-fold reduction in per-sample storage over raw features as a practical side benefit.
Abstract Quantum kernel approaches have been touted as a possible way to gain representational advantages over classical machine learning techniques however most comparative work only looks at classification accuracy leaving the geometric structure of the kernels unexplored. We carry out a comprehensive kernel-level comparison between classical support vector machines (SVMs) with radial basis function (RBF) kernels and quantum support vector classifiers that use ZZFeatureMap and PauliFeatureMap fidelity kernels, tested on two synthetic and two real-world two-dimensional datasets. After applying centred kernel alignment(CKA), the Hilbert-Schmidt Independence Criterion (HSIC), eigenvalue spectrum analysis, spectral entropy, and effective rank. Firstly, shallow quantum kernels form a moderately aligned band with RBF across all datasets and conditions that is quite stable (CKA : 0.41 → 0.57), whereas going deeper in the circuit to two repetitions largely mismatches both alignment and accuracy with no representational gain (e.g. Iris cross-validation accuracy: 0.920 → 0.670, p = 0.0004). Secondly, depolarising noise surprisingly increases the quantum-classical kernel alignment, which means that near-term hardware kernels are structurally closer to RBF than noiseless simulations suggest. Thirdly, the scenarios that lead to maximum apparent quantum-classical divergence, small sample sizes and low noise, are precisely those most used to benchmark quantum kernel advantage. Kernel-level structural analysis is essential for meaningful evaluation of quantum machine learning methods.