Quantum Kernel Advantage over Classical Collapse in Medical Foundation Model Embeddings Sebastian Cajas Ordóñez ,1, ∗ Felipe Ocampo Osorio ,1, 2 Dax Enshan Koh ,3, 4, 5 Rafi Al Attrach ,1 Aldo Marzullo ,6 Ariel Guerra-Adames ,7, 8 J. Alejandro Andrade ,9 Siong Thye Goh ,4, 10 Chi-Yu Chen ,11 Rahul Gorijavolu ,1, 12, 13, 14 Xue Yang ,15, 3, 16 Noah Dane Hebdon ,3 and Leo Anthony Celi 1, 17, 18
arXiv:2604.24597v1 [quant-ph] 27 Apr 2026
1
MIT Critical Data, Massachusetts Institute of Technology, Cambridge, MA, USA 2 Clinical Research Center, Artificial Intelligence Unit, Fundación Valle del Lili, Cali, Valle del Cauca, Colombia 3 Quantum Innovation Centre (Q.InC), Agency for Science, Technology and Research (A*STAR), 2 Fusionopolis Way, Innovis #08-03, Singapore 138634, Republic of Singapore 4 Institute of High Performance Computing (IHPC), Agency for Science, Technology and Research (A*STAR), 1 Fusionopolis Way, #16-16 Connexis, Singapore 138632, Republic of Singapore 5 Science, Mathematics and Technology Cluster, Singapore University of Technology and Design, 8 Somapah Road, Singapore 487372, Republic of Singapore 6 Department of Electronics, Information and Bioengineering, Politecnico di Milano, Milan, Italy 7 Bordeaux Population Health Research Center, Inserm U1219, Université de Bordeaux, F-33000, Bordeaux, France 8 Inria Bordeaux, Université de Bordeaux, F-33000 Bordeaux, France 9 Universidad del Cauca, Popayán, Colombia 10 Singapore Management University, 81 Victoria St, Singapore 188065 11 National Taiwan University Hospital 12 School of Medicine, Johns Hopkins University, Baltimore, MD, USA 13 Department of Biomedical Engineering, Johns Hopkins University, Baltimore, MD, USA 14 AI for Responsible, Generalizable, and Open Surgical (ARGOS) Research Group, Baltimore, MD, USA 15 School of Information Engineering, Shanghai Maritime University, Shanghai, 201306, China 16 Research Center of Intelligent Information Processing and Quantum Intelligent Computing, Shanghai, 201306, China 17 Laboratory for Computational Physiology, MIT, Cambridge, MA, USA 18 Department of Medicine, Beth Israel Deaconess Medical Center, Boston, MA, USA We provide evidence of quantum kernel advantage under noiseless simulation in binary insurance classification on MIMIC-CXR chest radiographs using quantum support vector machines (QSVM) with frozen embeddings from three medical foundation models (MedSigLIP-448, RAD-DINO, ViTpatch32). We propose a two-tier fair comparison framework in which both classifiers receive identical PCA-q features; at Tier 1 (untuned QSVM vs. untuned linear SVM, C = 1 both sides), QSVM wins minority-class F1 in all 18 tested configurations (10 embedding seeds; 17 at p < 0.001, 1 at p < 0.01, paired bootstrap). The classical linear kernel collapses to majority-class prediction (F1 = 0) on 90–100% of seeds at every qubit count, while QSVM maintains non-trivial recall. At q = 11 (MedSigLIP-448 plateau center), QSVM achieves mean F1 = 0.343 ± 0.170 vs. classical F1 = 0.050 ± 0.159 (∆F1 = +0.293, p < 0.001) without hyperparameter tuning. Under Tier 2 (untuned QSVM vs. C-tuned RBF SVM), QSVM wins all seven tested configurations (mean gain +0.068, max +0.112). Eigenspectrum analysis reveals the mechanism: multi-seed mean quantum kernel effective rank reaches 69.80 at q = 11, far exceeding the linear kernel rank of exactly q = 11, while classical collapse remains C-invariant. At q = 16, any concentration collapse is seed-dependent: multi-seed mean F1 is 0.377, a Tier-1 win. A full qubit sweep reveals architecture-dependent concentration onset across models. Code: https://github.com/sebasmos/qml-medimage.
I.
INTRODUCTION
Quantum machine learning (QML) promises computational advantages through the use of quantum feature maps that embed classical data into exponentially large Hilbert spaces [1, 2]. Quantum kernel methods, and in particular the quantum support vector machine (QSVM), realize this promise by computing inner products of quantum states instead of explicit feature vectors, potentially enabling richer decision boundaries with fewer param-
eters than classical alternatives [3, 4]. Despite considerable theoretical interest, empirical demonstrations of quantum advantage on real-world medical imaging tasks remain rare, partly because rigorous fair comparisons require careful control of hyperparameters, dimensionality, and regularization on both the classical and quantum sides [5, 6]. This work builds on preliminary results presented at the AIQxQIA 2025 workshop [7], substantially expanding the experimental scope with a two-tier fair comparison framework, multi-model evaluation, and mechanistic analysis of the classical kernel collapse phenomenon. We study binary insurance classification on the MIMIC-CXR chest radiograph dataset [8, 9], predicting whether a patient holds
2 Private insurance versus Medicaid/Medicare coverage. Recent work has shown that deep learning models trained on chest radiographs can predict attributes not visually apparent to clinicians, including self-reported race [10] and insurance type [11], even when images are clinically normal. This phenomenon has been hypothesized to arise from spurious correlations: image features statistically associated with demographic or socioeconomic variables, specific positioning conventions, or markers of cumulative environmental exposure, without any direct causal relationship to pathology [12]. When models capture these latent signals rather than genuine clinical features, performance becomes brittle outside the training distribution and errors concentrate in underrepresented groups [13]. Our objective is to evaluate whether quantum kernels improve separability within this representation space, without claiming the learned signal is clinically causal. These findings carry direct implications for health equity: if socioeconomic and demographic signals are encoded in medical images, clinical AI systems trained on those images risk learning and perpetuating disparities, a concern supported by evidence that chest X-ray classifiers systematically underdiagnose underserved populations [13, 14]. Insurance status is already recorded in the electronic health record, so the goal here is not clinical deployment. Insurance prediction provides a clinically grounded test of whether quantum feature maps can extract discriminative structure that classical kernels miss, on a task whose difficulty (subtle, distributed signal in a class-imbalanced setting) makes the comparison meaningful. Instead of hand-crafted image features, we extract highdimensional embeddings from three frozen medical foundation models (MedSigLIP-448 [15], RAD-DINO [16], ViT-patch32 [17]), compress them to q dimensions via PCA, and compare QSVM against classical SVM baselines at identical feature dimensionality. This setting is representative of realistic small-sample quantum pipelines: the quantum hardware constraint limits practical training to ∼2,000 samples, which is naturally met by PCA reduction to q ≤ 16 dimensions. This paper makes four contributions. 1. Quantum kernel advantage across all tested configurations. QSVM (C=1, reps=1 [§III C], trace normalization) beats an equally untuned linear SVM on minority-class F1 at all 18 model×qubit configurations (q ∈ {4, 6, 8, 9, 10, 11, 12, 16}, three models), validated across 10 embedding seeds (17 at p < 0.001, 1 at p < 0.01; paired bootstrap). Classical linear SVM collapses to F1 = 0 on 90– 100% of seeds at every qubit count. Against the best C-tuned RBF kernel at equal PCA dimensionality, QSVM still wins all 7 configurations (mean gain +0.068). 2. Structural explanation for the classical collapse. PCA-q compression leaves the linear kernel with effective rank equal to q (3.77–5.85 out of N = 1,896 training samples), making collapse
independent of regularization parameter C. The quantum kernel reaches 6.86 and 13.94 at q = 4 and q = 6 (seed 0; 1.82× and 2.52× the linear values), with the ratio growing with qubit count. A 10-seed rank-matched RBF experiment confirms the advantage extends beyond effective rank: QSVM outperforms an RBF kernel tuned to the same rank at all four qubit counts tested. 3. Three design rules for quantum kernel pipelines. Trace normalization is necessary for nonzero QSVM F1; Frobenius normalization collapses it to zero on all models. 1-DOF angle encoding (one Ry per qubit) consistently outperforms the 3-DOF variant (Rz-Ry-Rz). Increasing re-uploading depth at q = 8 degrades performance; the bottleneck is sample size rather than circuit capacity. 4. Architecture-dependent concentration. A sweep over q ∈ {2, . . . , 16} reveals model-specific behaviour: on seed 0, MedSigLIP-448 peaks at q = 11 then collapses at q = 16 (multi-seed mean 0.377, a Tier-1 win), while RAD-DINO and ViT-patch32 improve monotonically. The variation is consistent with data-dependent concentration rates described by Thanasilp et al. [18] and extends those findings to frozen medical foundation model embeddings. The paper is organized as follows. Related work (§ II) covers quantum kernel methods, medical foundation models, and quantum advantage benchmarking. The methods (§ III) cover the dataset, preprocessing, circuit design, kernel computation, and fair comparison framework. Results (§ IV) report the main experiments; ablations (§ V) address normalization, qubit count, circuit depth, and data-type variants. The discussion (§ VI) interprets the structural collapse mechanism and limitations; § VII concludes.
II.
RELATED WORK
Quantum kernel methods exploit the ability of quantum circuits to efficiently compute inner products in exponentially large feature spaces. Havlı́ček et al. [1] introduced the quantum kernel estimator and demonstrated that a quantum feature map ϕ(x) can produce kernels that are classically intractable to simulate and may offer a path to quantum advantage. Schuld and Killoran [2] showed that quantum models are equivalent to kernel methods with a specific quantum kernel, which unifies the circuit-based and kernel-based views of QML. Schuld [3] further clarified the connection between quantum models and kernel methods in the NISQ era. Liu et al. [4] provided a rigorous quantum advantage proof for specific classification problems, while Huang et al. [19] introduced the notion of quantum kernel alignment and showed that the quantum advantage is dataset-dependent. Kübler et al. [20] studied the geometric difference between quantum and
3 classical kernel matrices and identified conditions under which quantum kernels cannot outperform classical ones. Thanasilp et al. [18] and Larocca et al. [21] analyzed exponential concentration (barren plateaus in kernels) and provided theoretical motivation for why high-qubit quantum kernels can collapse. Abbas et al. [22] studied the effective dimension of quantum models; their connection between circuit expressivity and generalization parallels our effective-rank analysis of the kernel matrix. Collectively, these results establish that the theoretical promise of quantum kernels is real, but empirical demonstrations on clinical data remain rare. Foundation models pre-trained on large corpora of medical images provide rich, transferable representations that outperform task-specific models on downstream clinical tasks [15, 16]. RAD-DINO [16] is a vision transformer pre-trained on radiology images using self-supervised DINO objectives and produces 768-dimensional embeddings that capture anatomical structure. MedSigLIP448 [15] adapts the SigLIP vision-language pre-training to medical imaging at 448-pixel resolution and produces 448dimensional embeddings optimized for semantic similarity. ViT-patch32 [17] is a general-purpose vision transformer (patch size 32) that serves as a non-medical baseline embedding model. Freezing these models and using only their CLS-token embeddings as input features eliminates any confounds from fine-tuning. PCA reduction to q ≤ 16 dimensions brings the embedding dimensionality into alignment with current quantum hardware constraints naturally, without requiring heuristic truncation. Establishing rigorous quantum advantage is non-trivial. Jerbi et al. [5] surveyed quantum machine learning benchmarks and argued that classical baselines must be evaluated at equal computational resources to avoid inflated quantum advantage claims. Bowles et al. [6] demonstrated that many purported QML advantages vanish under fair classical comparisons. Peral-Garcı́a et al. [23] provide a comprehensive survey of QML applications that contextualizes our medical imaging use case within prior QML work. Our two-tier fair comparison framework is designed to address all of these methodological concerns. Despite these theoretical and methodological advances, most prior QML studies report results on synthetic data or small toy benchmarks. Havlı́ček et al. [1] demonstrated quantum kernel advantage on a synthetic 2D classification task but did not evaluate on real-world medical data. Liu et al. [4] proved a rigorous quantum speedup for specific engineered data distributions; however, their construction does not transfer directly to natural datasets. Bowles et al. [6] benchmarked QML models on over 160 tabular datasets and found quantum kernels competitive but rarely superior to classical methods when applied to raw features. Our work differs in a key respect: we classify frozen foundation-model embeddings rather than raw input features, which may provide a more favourable inductive bias for quantum kernels. Senokosov et al. [24] surveyed QML for medical imaging and noted that nearly all prior studies operate on small subsets of
standard datasets (e.g., 100–500 samples from MNIST or dermoscopy collections). To our knowledge, our 2,371sample MIMIC-CXR experiment is one of the larger real clinical imaging datasets on which QML has been evaluated; most published QML medical imaging experiments operate on 100–500 samples [24]. To our knowledge, no prior work has applied quantum kernel methods to insurance or social determinant classification from medical imaging data. The 18/18 Tier-1 win rate (multi-seed), the mechanistic explanation of classical kernel collapse via effective rank (Section IV C), and the scale of the clinical dataset distinguish this work from prior empirical QML studies that report marginal or inconsistent advantages on toy problems. The low-rank structure of classical kernel matrices clarifies when the quantum advantage window opens. Support vector machines [25, 26] classify data by finding a maximum-margin hyperplane in feature space. The kernel trick enables non-linear classification by implicitly mapping inputs to a reproducing kernel Hilbert space (RKHS). The effectiveness of any kernel depends critically on the rank structure of the resulting kernel matrix: a low-rank kernel matrix cannot distinguish samples whose projections onto the kernel’s feature space coincide. This observation forms the theoretical basis for understanding classical collapse at low PCA dimensionality (Section IV C).
III. A.
METHODS
Dataset and Task
The MIMIC-CXR dataset [8] contains de-identified chest radiographs from approximately 61,000 patients with associated clinical metadata. Johnson et al. [9] released the JPEG version (MIMIC-CXR-JPG) with structured labels derived from free-text radiology reports. Insurance type is recorded in the hospital admission record linked to each study and enables the insurance classification task studied here. We use the MIMIC-CXR-JPG dataset [9] restricted to the DT9 preprocessing stratum, which enforces one image per patient (preventing data leakage from repeated studies), removes duplicate filenames, and retains only samples with valid binary insurance labels. DT9 corresponds to the “Uncertainty Coreset” preprocessing stratum [27], which selects one image per patient via coreset sampling, removes duplicate filenames, and retains only samples with valid binary insurance labels (Medicare/Medicaid vs. Private). The result is N = 2,371 samples. This stratum was selected because it produced the strongest quantum results in preliminary experiments, which constitutes a post-hoc choice. Two observations mitigate the resulting multiplecomparisons concern: (1) the classical kernel collapse is structural (effective rank = q) and occurs across all strata, so the collapse-regime wins are not DT9-specific;
4 (2) the preprocessing pipeline (StandardScaler→PCAq→MinMaxScaler[−1,1]) is identical across all strata, so the quantum circuit sees identically scaled inputs regardless of stratum. The non-collapse Tier-1 advantage (q ≥ 10) has been validated only on DT9; confirming it on additional strata remains future work. The classification target is binary: Medicaid/Medicare patients are assigned class 0 (majority, 69.6%) and Private insurance patients are assigned class 1 (minority, 30.4%). The resulting dataset contains approximately Ntotal ≈ 2,371 samples split into training (Ntrain = 1,896), validation, and test sets using a fixed random seed (seed 0) with an 80/10/10 ratio. The strong class imbalance (69.6%/30.4%) means that a majority-class predictor achieves accuracy ≈ 0.697 but minority-class F1 = 0. Following Sokolova and Lapalme [28], who recommend class-aware metrics for imbalanced binary classification, we report minority-class F1 (i.e., F1 for the positive class, Private insurance) as the primary evaluation metric; accuracy and AUC are reported as secondary metrics. B.
Embeddings and Preprocessing
We extract frozen embeddings from three publicly available foundation models: 1. MedSigLIP-448: 448-dimensional CLS-token embeddings from a medical SigLIP model fine-tuned at 448-pixel resolution. 2. RAD-DINO: 768-dimensional CLS-token embeddings from a DINO self-supervised vision transformer pre-trained on radiology images. 3. ViT-patch32-cls: 768-dimensional CLS-token embeddings from a general-purpose ViT with patch size 32 (no domain-specific pre-training). We use CLS-token pooling as the primary embedding strategy for all models. A global average pooling (GAP) variant of ViT-patch32 (ViT-patch32-GAP, 768dimensional) was also evaluated across 10 seeds as a pooling ablation; results are reported in Appendix A 5. All embeddings are processed through the same threestage pipeline (Fig. 1). StandardScaler normalizes to zero mean and unit variance; PCA reduces to q dimensions and retains 5.6%–41.1% of explained variance depending on model and q (Table V); MinMaxScaler re-scales to [−1, 1] to match the angle encoding range of the quantum circuit. We tested q ∈ {2, 3, 4, 5, 6, 8, 9, 10, 11, 12, 16} qubits across experiments (see Figure 11 in the Appendix for a 2D PCA scatter illustrating the low explained variance at q = 2). C.
Quantum Circuit and Kernel
We adopt the Block-Sparse Parameterization (BSP) circuit with one degree of freedom (1-DOF) per qubit:
Raw
Std
PCA-q
MinMax
x ∈ RD
µ=0, σ=1
RD → Rq
[−1, 1]q
FIG. 1. Three-stage preprocessing pipeline applied to all embeddings. D ∈ {448, 768} depending on the foundation model; q ∈ {2, 3, 4, 5, 6, 8, 9, 10, 11, 12, 16} is the qubit/feature count.
each qubit receives a single parameterized Ry rotation encoding one PCA component. The circuit structure for q qubits is: q h i Y U (u) = CNOTd, (d mod q)+1 · Ry (ud ) ,
(1)
d=1
where the subscript d mod q + 1 implements ring entanglement: qubit q connects back to qubit 1. The number of times this encoding block is applied sequentially is referred to as the data re-uploading depth (reps); each repetition re-encodes the full input vector into the circuit. The data re-uploading depth is fixed at reps = 1 for all primary experiments. The quantum kernel is computed via the compute– uncompute strategy: 2
KQ (ui , uj ) = ⟨0q |U † (ui )U (uj )|0q ⟩ .
(2)
Trace normalization. We apply trace normalization to the raw kernel matrix before passing it to the SVM solver: K̃Q =
KQ , tr(KQ )
(3)
which sets tr(K̃Q ) = 1 and makes kernels of different scale comparable before the SVM solver. The testtrain kernel block is normalized by the same training trace: K̃Q,test = KQ,test /tr(KQ ). Section V shows that trace normalization is critical: Frobenius normalization collapses F1 to zero across all models. Kernel computation uses Qiskit 1.2.4’s Statevector simulator with the renew operand fast-path (a public QuantumCircuit method introduced in Qiskit 1.x for in-place operand reuse) for efficient batch evaluation.
D.
Fair Comparison Framework
A methodologically rigorous quantum advantage claim requires careful control of all hyperparameters on both sides. We define a two-tier framework: Tier 1: The fair fight.: Untuned QSVM (C = 1, q qubits) vs. untuned linear SVM (C = 1, PCA-q features). Both sides have the same regularization, same dimensionality, and neither is cross-validated. We emphasize that identical C does not imply
5 TABLE I. Two-tier comparison framework. Tier 1 is the primary claim; Tier 2 tests against a tuned classical baseline. Tier QSVM
Classical SVM
Result
1
C=1, q qubits C=1, PCA-q 18/18 F1 wins (untuned) linear (untuned) (10 seeds, paired bootstrap)
2
C=1, q qubits best-C, PCA-q (untuned) rbf (C-tuned)
7/7 F1 wins avg +0.068 F1
master reproduction script (scripts/run all.sh) executes the full experimental pipeline, from embedding loading through figure and table generation. Single-seed results can be reproduced in approximately 12 GPU-hours on an NVIDIA H100 (full qubit sweep for one model); classical baselines require only CPU and complete in minutes.
IV.
equivalent regularization across kernels. Tier-1 isolates performance under identical hyperparameter choices, not identical effective capacity. Tier-1 is the primary paper claim. Tier 2: The stretched goal.: Untuned QSVM (C = 1, q qubits) vs. C-tuned RBF (γ at sklearn’s default scale, C over {0.01, 0.1, 1, 10, 100}) at the same PCAq dimensionality. Winning Tier 2 on F1 despite the classical side having hyperparameter tuning is strong evidence of genuine quantum advantage. Table I summarizes the two-tier framework.
E.
Evaluation Metrics
We report test-set accuracy, minority-class F1 score, and AUC-ROC. Given the 70/30 class imbalance, minority-class F1 is the primary metric: it directly measures whether the classifier detects the underrepresented group, penalizing majority-class collapse in a way that accuracy does not [28–30]. A classifier that never identifies the minority class is clinically useless regardless of its overall accuracy. In this task, the minority class (Private insurance, 30.4%) is the group whose misclassification carries downstream resource-allocation and health-equity consequences; recall on that class, not aggregate accuracy, is the operationally meaningful quantity. Multi-seed statistical validation (10 seeds × 5 models × 11 qubit counts = 550 QSVM configs, plus 550 × 2 classical SVM configs covering linear and tuned kernels at matching PCA dimensions = 1,100 classical configs) confirms the advantage: QSVM wins all 18 Tier-1 configurations on mean F1 (17 at p < 0.001, 1 at p < 0.01; paired bootstrap, 10,000 resamples, seed 42). Classical linear SVM collapses to F1 = 0 on 90–100% of seeds at every qubit count tested (Table II; full breakdown in Section IV A).
F.
Reproducibility
All source code, SLURM job configurations, and analysis scripts are available at https://github. com/sebasmos/qml-medimage. Pre-computed foundation model embeddings (20 seeds per model) are hosted at https://huggingface.co/datasets/ MITCriticalData/qml-mimic-cxr-embeddings. A
A.
RESULTS
Tier 1: Fair Comparison
Table II reports the primary Tier-1 comparison across all three models at all measured PCA-q dimensions. Multiseed validation (10 independent embedding seeds) reveals that classical collapse is pervasive: classical linear SVM collapses to F1 = 0 on 100% of seeds at q ≤ 9 (MedSigLIP), q ≤ 10 (RAD-DINO), and all tested q (ViTpatch32). Even at higher qubit counts where seed 0 showed a working classical baseline, 9 of 10 seeds collapse. The “non-collapse” regime observed in single-seed analysis was a seed 0 artefact. QSVM (C = 1, trace normalization, reps = 1) wins on mean minority-class F1 in all 18 configurations (Table II; 17 at p < 0.001, 1 at p < 0.01, paired bootstrap). The classical collapse is C-invariant: re-running with C ∈ {0.01, 0.1, 1, 10, 100} produces the same majority-class prediction (F1 = 0) at PCA-4 and PCA-6 for all three models. Classical baselines extended to all qubit counts. Relative to the six low-q configurations reported in our preliminary work [7], we now provide classical SVM (C = 1) baselines at all PCA-q dimensions matching our QSVM qubit counts (Table II). QSVM wins on minority-class F1 in all 18 configurations across 10 embedding seeds. The strongest single-configuration result is MedSigLIP448 at q = 11: mean QSVM F1 = 0.343 ± 0.170 versus classical F1 = 0.050 ± 0.159 (∆F1 = +0.293, 95% CI [+0.190, +0.385], p < 0.001). For RAD-DINO specifically, the quantum advantage extends to accuracy as a secondary metric: QSVM achieves statistically significant accuracy gains at q ∈ {4, 6, 8, 10} (∆acc = +1.2–1.8%, p ≤ 0.005, paired bootstrap); the advantage is not solely an artifact of minority-class F1 sensitivity under class imbalance. Strongest configuration: q = 11. At q = 11 qubits, MedSigLIP-448 QSVM achieves mean F1 = 0.343 ± 0.170 versus classical linear SVM F1 = 0.050 ± 0.159 (∆F1 = +0.293, 95% CI [+0.190, +0.385], p < 0.001, 10 seeds). Both classifiers use C = 1 (no tuning), ruling out hyperparameter cherry-picking. Classical linear SVM collapses to F1 = 0 on 9 of 10 seeds at q = 11; QSVM maintains F1 > 0.1 on 8 of 10 seeds, remaining non-trivial across the full seed distribution. Table III shows the confusion matrix for a representative QSVM run at q = 11 (seed 0). The minority class (Private insurance) achieves precision 0.639 and recall 0.542; the F1 = 0.586 reflects balanced minority-class
6 TABLE II. Extended Tier-1 comparison: classical linear SVM C = 1 vs. QSVM C = 1 at all measured PCA-q dimensions (DT9, mean ± std over 10 embedding seeds). Bold indicates the Tier-1 winner on mean F1. Model
q
QSVM C=1 Acc
F1
Acc
F1
4 6 8 9 10 11 12 16
0.697±0.016 0.698±0.026 0.702±0.025 0.699±0.020 0.704±0.032 0.704±0.027 0.704±0.030 0.700±0.024
0.212±0.157 0.286±0.156 0.323±0.163 0.333±0.168 0.353±0.173 0.343±0.170 0.374±0.146 0.377±0.147
0.696±0.004 0.696±0.004 0.696±0.004 0.696±0.004 0.702±0.021 0.702±0.021 0.703±0.022 0.703±0.023
0.000±0.000 0.000±0.000 0.000±0.000 0.000±0.000 0.050±0.159 0.050±0.159 0.051±0.161 0.050±0.160
F1 WIN F1 WIN F1 WIN F1 WIN F1 WIN F1 WIN F1 WIN F1 WIN
RAD-DINO RAD-DINO RAD-DINO RAD-DINO RAD-DINO
4 6 8 10 16
0.713±0.020 0.713±0.023 0.714±0.019 0.708±0.014 0.712±0.019
0.332±0.184 0.371±0.143 0.400±0.120 0.406±0.114 0.450±0.083
0.696±0.004 0.696±0.004 0.696±0.004 0.696±0.004 0.702±0.012
0.000±0.000 0.000±0.000 0.000±0.000 0.000±0.000 0.079±0.127
F1 WIN F1 WIN F1 WIN F1 WIN F1 WIN
ViT-p32 ViT-p32 ViT-p32 ViT-p32 ViT-p32
4 6 8 10 16
0.696±0.007 0.698±0.021 0.703±0.024 0.707±0.029 0.702±0.033
0.048±0.073 0.272±0.132 0.370±0.092 0.391±0.081 0.399±0.098
0.696±0.004 0.696±0.004 0.696±0.004 0.696±0.004 0.696±0.004
0.000±0.000 0.000±0.000 0.000±0.000 0.000±0.000 0.000±0.000
F1 WIN F1 WIN F1 WIN F1 WIN F1 WIN
Pred. Medicare Pred. Private True Medicare True Private
144 33
Table IV shows the Tier-2 comparison: untuned QSVM (C=1) vs. C-tuned RBF (γ at sklearn default scale) at the same PCA dimensionality, reported as mean ± std over 10 embedding seeds. QSVM wins minority-class F1 on all seven configurations across seeds. Figure 5 shows the qubit scaling behavior across all three embedding models.
Classical Kernel Collapse Analysis
The root cause of the classical collapse is consistent with structural rank limitations rather than a tuning artifact. After PCA reduction to q dimensions, the linear kernel matrix KL = Xnorm X⊤ norm has at most q non-zero eigenvalues out of N = 1,896 training samples. Table V reports the effective rank [20]
i
λi pi = P , j λj
q
QSVM C=1 Acc
Tier 2: Clinical F1 Advantage
X eff rank = exp − pi log pi ,
TABLE IV. Tier-2 clinical F1 advantage: QSVM (C=1, trace) vs. best rbf SVM (C-tuned, ∗ = C-tuned for MedSigLIP) at equal PCA-q dimensionality. Mean ± std over 10 embedding seeds; ∆F1 is mean QSVM F1 minus mean best-classical F1. Model
22 39
detection rather than a precision-recall trade-off artifact.
C.
Verdict
MedSigLIP MedSigLIP MedSigLIP MedSigLIP MedSigLIP MedSigLIP MedSigLIP MedSigLIP
TABLE III. Confusion matrix for MedSigLIP-448 QSVM at q = 11 (C=1, trace normalization, DT9, seed 0). N = 238 test samples.
B.
Lin. SVM C=1
(4)
F1
Best SVM Kernel
∆F1
Rel. Gain
F1
MedSigLIP 4 0.697±0.016 0.212±0.157 MedSigLIP 6 0.698±0.026 0.286±0.156 MedSigLIP 8 0.702±0.025 0.323±0.163
rbf∗ rbf∗ rbf∗
0.169±0.097 +0.043 0.178±0.123 +0.109 0.211±0.126 +0.112
+26% +61% +53%
RAD-DINO 4 0.713±0.020 0.332±0.184 RAD-DINO 6 0.713±0.023 0.371±0.143
rbf rbf
0.264±0.133 +0.069 0.285±0.096 +0.086
+26% +30%
ViT-p32 4 0.696±0.007 0.048±0.073 rbf 0.044±0.043 +0.004 ViT-p32 6 0.698±0.021 0.272±0.132 rbf 0.219±0.091 +0.054 ∗ C-tuned: grid search over C ∈ {0.01, 0.1, 1, 10, 100}.
+9% +25%
which quantifies how uniformly eigenvalue mass is distributed. With effective rank ≈ q ≪ N , virtually all 1,896 training samples project onto the same q-dimensional subspace. Classical SVM with a linear kernel operating on these degenerate representations fails to separate the minority class in practice, often defaulting to majority-class prediction, regardless of the regularization parameter C. Collapse statistics (Table VI) confirm that the intra-class kernel variance equals the total variance (within-class KL similarity ≈ between-class), which leaves little to no usable discriminative signal. The quantum feature map U (u) implicitly operates in a 2q -dimensional Hilbert space (16 or 64 dimensions for q = 4 or q = 6), which is exponentially larger than the qdimensional subspace accessible to the linear kernel [1, 2]. This expanded feature space is reflected directly in the measured kernel matrices: for MedSigLIP-448 (seed 0),
7 TABLE V. Kernel effective rank at PCA-q dimensionality. N = 1,896 training samples. For q ≤ 6, the linear kernel KL has exactly q positive eigenvalues (Shannon eff. rank: 3.77– 5.85 out of 1896), which is consistent with structural collapse. † Quantum kernel KQ effective rank; non-dagger rows report KL (linear kernel) statistics. PCA-11 classical linear KL (seed 0): acc = 0.761, F1 = 0.504 (seed 0), eff. rank ≈ 10.2 (noncollapse; QSVM wins by +0.082 F1; ratio 43.04/10.2 ≈ 4.2× cited in the abstract). At q = 16, PCA var%, Nλ>0 , and λmax are not reported because swap-test fidelity concentration causes all off-diagonal kernel entries to converge toward a single value, which renders within-class and between-class variance statistics uninformative; the effective rank (92.13) is retained as it directly quantifies the concentration onset (see Section V E).
Model
q
MedSigLIP MedSigLIP MedSigLIP† MedSigLIP† RAD-DINO RAD-DINO ViT-p32 ViT-p32
4 6 11 16 4 6 4 6
PCA var% Nλ>0 Eff. Rank λmax 32.6 41.1 56.0 — 5.6 7.7 28.4 34.7
4 6 — — 4 6 4 6
3.77 5.53 43.04 92.13 3.89 5.85 3.86 5.59
770.6 614.8 468.6 — 666.1 475.5 627.5 502.3
TABLE VI. Linear kernel KL variance statistics (200 subsampled training samples, sorted by class). Low variance and within-class ≈ between-class means no discriminative signal for the classical SVM. Model
q KL mean KL std KL var
MedSigLIP MedSigLIP RAD-DINO RAD-DINO ViT-p32 ViT-p32
4 6 4 6 4 6
0.0728 0.1049 0.1387 0.1263 0.2222 0.2022
0.5099 0.4181 0.4944 0.4097 0.4657 0.3893
0.2600 0.1748 0.2444 0.1678 0.2169 0.1516
KQ achieves Shannon effective ranks of 6.86 at q = 4 and 13.94 at q = 6, vs. 3.77 and 5.53 for KL (1.82× and 2.52× the linear values; Figure 2). The ratio grows with qubit count, consistent with the quantum feature map accessing an exponentially larger Hilbert space (Figure 3). This directly measured higher effective rank of KQ is the structural explanation for why QSVM maintains non-trivial F1 in the same PCA subspace where classical kernels collapse. At the performance peak (q = 11), the seed 0 quantum kernel effective rank reaches 43.04, a 6.3× increase over q = 4 (multi-seed mean: 69.80). The optimal qubit count coincides with maximal kernel expressiveness.
D.
FIG. 2. Linear kernel KL eigenspectrum for MedSigLIP-448 at q = 6. The kernel has exactly 6 positive eigenvalues (effective rank 5.53) out of N = 1,896 training samples, which confirms that PCA-6 compression collapses the kernel matrix to a 6dimensional subspace.
Feature Selection Sensitivity
To verify that the quantum advantage is not an artifact of PCA-based dimensionality reduction, we replaced PCA with two alternative feature selection methods: mutual
FIG. 3. Quantum vs. linear kernel eigenspectrum comparison for MedSigLIP-448 at q ∈ {4, 6} (left: normalized eigenvalue decay; right: cumulative energy). The quantum kernel KQ has 1,133 (q=4) and 1,586 (q=6) eigenvalues exceeding numerical threshold ε = 10−10 ; the theoretical upper bound for the fidelity kernel is 4q (256 and 4,096 for q = 4, 6), so the surplus above that bound consists of finite-precision numerical artefacts from statevector simulation. For comparison, the linear kernel has exactly 4 and 6 algebraically positive eigenvalues. The red vertical line marks the rank boundary of the linear kernel (Nλ>0 = q), beyond which all linear kernel eigenvalues collapse to numerical zero; the quantum kernel eigenspectrum extends far beyond this boundary, a direct consequence of its higher effective rank. Shannon effective rank (seed 0): quantum q = 4: 6.86 (1.82× linear 3.77); quantum q = 6: 13.94 (2.52× linear 5.53). The ratio grows with qubit count, consistent with the quantum feature map accessing an exponentially larger Hilbert space (2q dims).
information (MI) ranking and kernel PCA (kPCA), each selecting k ∈ {4, 6} features. The best classical SVM F1 with optimal MI/kPCA feature selection was: MedSigLIP448, 0.404; RAD-DINO, 0.186; ViT-patch32, 0.267. All three remain below the corresponding QSVM F1 at q = 4 (0.488, 0.448, 0.184) and q = 6 (0.504, 0.435, 0.422). The quantum advantage holds across all three dimensionality reduction methods and is not an artifact of PCA geometry.
8 TABLE VII. Effect of kernel normalization on QSVM (q = 8, reps=1, C=1, DT9, seed 0). Trace normalization is critical for non-zero F1. Model
Norm.
Acc
AUC
F1
MedSigLIP MedSigLIP MedSigLIP MedSigLIP
trace none cosine frobenius
0.756 0.756 0.756 0.697
0.686 0.690 0.690 0.705
0.554 0.420 0.420 0.000
RAD-DINO RAD-DINO RAD-DINO RAD-DINO
trace none cosine frobenius
0.734 0.738 0.730 0.696
0.631 0.661 0.656 0.655
0.496 0.354 0.360 0.000
ViT-p32 ViT-p32 ViT-p32 ViT-p32
trace none cosine frobenius
0.744 0.735 0.735 0.697
0.652 0.658 0.658 0.671
0.450 0.337 0.337 0.000
FIG. 4. Quantum kernel matrix KQ (trace-normalized) for MedSigLIP-448 at q = 6 (200 training samples sorted by class label). The off-diagonal block structure reflects class boundaries; the quantum feature map preserves discriminative signal in the same PCA subspace where the linear kernel collapses (Table VI).
V.
ABLATION STUDIES
A.
Kernel Normalization
Table VII reports the effect of four kernel normalization strategies on the QSVM with q = 8 qubits. Trace normalization achieves the best F1 (0.554 for MedSigLIP). Frobenius normalization collapses F1 to 0 on all three models, the same failure mode as the linear SVM, consistent with Thanasilp et al. [18] who showed that global rescaling can destroy the discriminative information in a quantum kernel. Unnormalized (none) and cosine normalization are intermediate: they match each other on accuracy but have lower F1 than trace. We replicate this finding at q = 2 and q = 3 for RAD-DINO, where all three non-trace normalizations collapse to F1 = 0 at q = 2, and cosine/none reach only F1 = 0.054 at q = 3. These results confirm that trace normalization is optimal across the full qubit range tested. B.
Qubit Count and Data Re-Uploading Depth
Multi-seed results (Table II) show QSVM winning in all 18 configurations across all three models. For MedSigLIP-448, the q = 9–12 window forms a clean Tier-1 plateau: multi-seed QSVM outperforms classical at every qubit count, with all four configurations using C = 1 and exceeding the PCA-matched classical ceiling by +0.052 to +0.579. At q = 16, multi-seed mean QSVM F1 is 0.377, a Tier-1 win. The partial seed 0 qubit sweep (q ∈ {2, 3, 4, 5, 6, 8}, Figure 5) illustrates the
FIG. 5. Partial qubit sweep (q ∈ {2, 3, 4, 5, 6, 8}, C = 1, DT9, seed 0): QSVM test accuracy (left) and minority-class F1 (right) for all three models. MedSigLIP-448 reaches F1 = 0.554 at q = 8; RAD-DINO and ViT-p32 improve monotonically across this range with ViT collapsing at q = 3 (F1 = 0). Full results at q ∈ {9, 10, 11, 12, 16} are reported in Table II.
non-monotonic shape: MedSigLIP-448 shows a plateau from q = 9 to q = 12 (seed 0 F1: 0.552, 0.578, 0.586, 0.561), then drops to F1 = 0.173 at q = 16 on seed 0, while the multi-seed mean (0.377) remains a Tier-1 win — the collapse at q = 16 is seed-dependent, not structural. RAD-DINO and ViT-patch32-cls improve F1 more monotonically up to q = 16 (F1 = 0.524 and 0.520 respectively); at q = 10, RAD-DINO reaches F1 = 0.488 and ViT reaches F1 = 0.478, both continuing their monotonic rise. The per-model variation suggests that barren-plateau concentration [18, 21] is embedding-specific at q = 16: MedSigLIP-448 embeddings are more susceptible to kernel concentration at high qubit counts than RAD-DINO or ViT-patch32. This is consistent with Thanasilp et al. [18], who showed that concentration rates depend on data structure and circuit architecture. Increasing the data re-uploading depth (reps) from 1 to 2 at q = 8 degrades performance: reps = 2 gives acc = 0.727 (vs. reps = 1: acc = 0.756), a drop of −0.029. Reps = 3 was cancelled. This further supports the nonmonotonic qubit curve interpretation: more expressive circuits do not reliably improve performance at this sample size.
9 TABLE VIII. 1-DOF vs. 3-DOF circuit comparison (q = 8, reps=1, trace normalization, C=1, DT9, seed 0). 3-DOF uniformly collapses. Model
Circuit
Acc
AUC
F1
MedSigLIP MedSigLIP
1-DOF 0.752 0.687 0.543 3-DOF 0.328 0.668 0.286
RAD-DINO 1-DOF 0.743 0.636 0.496 RAD-DINO 3-DOF 0.333 0.552 0.392 ViT-p32 ViT-p32
1-DOF 0.735 0.650 0.388 3-DOF 0.387 0.669 0.198
TABLE IX. q = 16 C-tuning results (DT9, seed 0). BestC selected by validation F1. Compare to q = 8 (best-C): MedSigLIP F1=0.554, RAD-DINO F1=0.507, ViT F1=0.446. Model
Best-C
MedSigLIP-448 RAD-DINO ViT-patch32-cls
C.
0.1 1.0 1.0
Acc
F1
vs q8 ∆F1
0.718 0.173 0.747 0.524 0.744 0.520
−0.381 +0.028 +0.074
Circuit Depth: 1-DOF vs. 3-DOF
We compared the 1-DOF circuit (one Ry parameter per qubit, q total parameters) against a 3-DOF variant (RzRy-Rz per qubit, 3q parameters) at q = 8. The 3-DOF circuit collapses on all three models: accuracy drops to 0.33– 0.39 (near random chance), F1 falls to 0.19–0.39. The 1-DOF circuit achieves acc = 0.735–0.756 and F1 = 0.388– 0.543 on the same datasets. Over-parameterization of the angle encoding circuit (3-DOF) appears to destroy the structured quantum interference that gives rise to the useful quantum kernel, consistent with the trainability arguments in the barren plateau literature (Table VIII) [31].
E.
Projected Quantum Kernel at q = 16
To determine whether the MedSigLIP-448 collapse at q = 16 is specific to the swap-test fidelity measurement or inherent to the quantum circuit structure, we ran the projected quantum kernel of Huang et al. [19] at q = 16 (DT9, seed 0). The projected kernel replaces the O(N 2 ) pairwise inner products with O(N ) Pauli-Z expectation values per sample and builds an RBF kernel on those expectation vectors. Grid search over γ ∈ {0.5, 1, 2, 5, 10} and C ∈ {0.01, 0.1, 1, 10, 100} selected γ = 5, C = 1 by validation accuracy (not F1; this is a mechanistic diagnostic, separate from the Tier-1/2 comparison). The projected kernel recovers minority-class F1 from 0.173 (fidelity q = 16, seed 0) to 0.396 (+0.223). This is a mechanistic diagnostic finding: the BSP circuit at q = 16 still encodes discriminative information: the seed 0 F1 recovery from 0.173 to 0.396 provides evidence for it. The bottleneck is the swap-test fidelity measurement, not the quantum feature map itself. The exponential concentration of |⟨ψx |ψy ⟩|2 at q = 16 destroys inter-sample contrast, while the projected kernel, operating on 16-dimensional Pauli-Z expectation vectors, is immune to this concentration by design. For MedSigLIP-448 at seed 0, this is consistent with a measurement bottleneck rather than a circuit expressibility limit — a mechanistic finding with direct implications for choosing kernel estimation methods in near-term QML: when fidelity-based kernels concentrate at high qubit counts, projected kernels provide a principled remedy. Whether this generalises across seeds and models remains an open question. The projected variant does not, however, surpass the fidelity peak at q = 11 (F1 = 0.586, seed 0); the optimal quantum advantage regime for MedSigLIP-448 remains q ≤ 11 under the 1-DOF BSP circuit.
VI. A.
D.
q16 Extended Results
Full C-tuning at q = 16 (three × H100, 400 GB host RAM each) reveals strongly model-dependent behaviour at q = 16 (Table IX). RAD-DINO achieves F1 = 0.524 at q = 16 (best-C = 1.0), which exceeds its q = 8 result (F1 = 0.496, +0.028); ViT-patch32-cls similarly improves to F1 = 0.520 (+0.074 vs. q = 8 F1 = 0.446). MedSigLIP448 collapses most severely on seed 0 (F1 = 0.173, bestC = 0.1), consistent with the original C = 1 result. The per-model variation (RAD-DINO and ViT improve F1 toward q = 16 while MedSigLIP-448 collapses) suggests model-specific concentration behaviour, and is captured in Table II (seed 0 partial sweep in Figure 5).
DISCUSSION
Why Classical Kernels Collapse
The classical collapse is a structural consequence of dimensionality, not a failure of hyperparameter tuning. After PCA-q reduction with q ≤ 6 (and in fact up to q = 9 for MedSigLIP, q = 10 for RAD-DINO, and q = 16 for ViTpatch32), the linear kernel KL lives in a q-dimensional subspace of a 1896-sample space. With effective rank ≈ q (Table V), the Gram matrix is essentially rank-q: nearly all pairs of training samples are mapped to identical points in the kernel’s implicit feature space. Classical SVM, which finds a maximum-margin hyperplane in this space, cannot distinguish minority-class samples from the majority class because their kernel representations are indistinguishable. The C-invariance of the collapse (tested over five decades: C from 0.01 to 100) confirms that no amount of regularization tuning can rescue a structurally degenerate kernel. The empirical measurements underpinning this argument (effective rank, variance de-
10 composition, and kernel heatmaps) are reported in § IV C (Tables V and VI, Figure 4). The quantum feature map U (u) maps q-dimensional inputs to a 2q -dimensional Hilbert space via entangling Ry rotations. The resulting quantum kernel KQ can have effective rank up to 4q (the operator-space dimension: 256 or 4,096 for q = 4 or q = 6), a qualitative difference from KL whose effective rank is exactly q. The empirical evidence, non-zero QSVM F1 while classical F1 is zero in the same PCA subspace, is consistent with this structural argument. One limitation of this explanation deserves mention: an RBF kernel with tuned bandwidth γ can achieve effective rank approaching N , far exceeding the quantum kernel’s 43.04 at q = 11 (seed 0). The Tier-2 results show QSVM beating tuned RBF kernels on F1, but that comparison tunes C only, not γ at fixed PCA-q. The quantum advantage may therefore reflect favorable inductive bias (the specific spectral structure of the quantum kernel) rather than rank alone. We test this directly via a 10-seed rank-matched RBF experiment (Table X, § VI F): γ ∗ is set per seed to match eff rank(KQ ). At q = 4, rank-matched RBF collapses on 30 % of seeds and achieves mean F1 = 0.110, compared with 20 % collapse and mean F1 = 0.212 for QSVM, despite identical effective rank. QSVM outperforms rank-matched RBF at all qubit counts, indicating that the quantum feature map’s spectral structure, beyond its effective rank, contributes to collapse resistance and predictive performance.
B.
The q=11 Tier-1 Win: Closing the Quantum Advantage Gap
The most significant configuration in this work is MedSigLIP-448 QSVM at q = 11. Across 10 embedding seeds, QSVM achieves mean F1 = 0.343 ± 0.170 vs. classical linear SVM F1 = 0.050 ± 0.159 (∆F1 = +0.293, 95% CI [+0.190, +0.385], p < 0.001, paired bootstrap). The seed 0 run (F1 = 0.586 vs. 0.504) is the hardest test for this comparison: it is the one seed where the classical SVM does not collapse and instead produces a valid, nontrivial classifier at PCA-11. On 9 of 10 seeds the classical baseline collapses to F1 = 0; QSVM surpasses it on all 10, without any hyperparameter tuning. Three aspects of this result merit discussion. First, both classifiers use C = 1 and receive identical PCA-11 features, so the accuracy and F1 gains (+0.008 and +0.082, respectively) are attributable solely to the quantum feature map. A caveat applies: C = 1 is a coincidentally reasonable default for the quantum kernel, whose higher effective rank provides a well-conditioned optimization surface, while for the collapsed classical kernel at low q the value of C is irrelevant (collapse is C-invariant). At q = 11, where the classical SVM is functional, this asymmetry is less pronounced and the comparison more defensible. The minority-class F1 advantage (+0.082) also carries clinical weight because it reflects improved detection of the minority class on a task with direct health-equity relevance.
Systematic misclassification of insurance status could obscure disparities in care between publicly and privately insured patients. Such underestimation of minority-class patients can propagate algorithmic bias with direct implications for health equity [14]. Beyond accuracy, the result closes the quantum advantage gap established at lower qubit counts: the BSP angle-encoding circuit avoids the classical collapse regime and exceeds the classical non-collapse ceiling at the right qubit count. Quantum advantage is achievable across both regimes within a single model and circuit family. The non-monotonic nature of the seed 0 qubit curve (plateau at q = 9–12 with peak at q = 11, seed 0 collapse at q = 16) further implies that qubit count is a tunable design variable; the quantum advantage window exists and can be identified by sweeping q. The entire q = 9–12 plateau forms a clean Tier-1 window (all use C = 1), with q = 9 achieving F1 = 0.552 (seed 0) while classical PCA-9 collapses to F1 = 0.
C.
Foundation Model Choice
MedSigLIP-448 consistently outperforms RAD-DINO and ViT-patch32 in the quantum setting across all qubit counts. Multi-seed mean F1 reaches 0.343 ± 0.170 at q = 11 and 0.377 at q = 16, both Tier-1 wins. Seed 0 peaks at F1 = 0.586 (q = 11), useful for circuit diagnostics but not the headline figure. RAD-DINO is second; ViT-patch32 (general domain) is weakest. This ordering mirrors the expected quality of medical domain alignment: MedSigLIP is trained explicitly for medical image-text alignment at high resolution, while ViT has no domainspecific pre-training. The result suggests that quantum kernels amplify the quality of the underlying embedding: better-aligned foundation models provide richer PCA subspaces that the quantum feature map can exploit. The concentration phenomenon is embedding-specific rather than a universal limitation of the BSP circuit: RAD-DINO and ViT-patch32-cls show monotonic F1 improvement from q = 2 through q = 16 (F1 = 0.176→0.524 and 0.104→0.520 respectively), with no peak or collapse. Only MedSigLIP-448 exhibits the non-monotonic peakthen-collapse pattern. This suggests that the concentration rate depends on the structure of the embedding space, not solely on circuit depth or qubit count, consistent with Thanasilp et al. [18], who showed that data distribution and encoding architecture jointly determine when exponential concentration sets in. The eigenspectrum progression provides a complete mechanistic narrative (see Figures 7 and 8 in the Appendix for the q = 11 eigenspectrum and kernel heatmap): seed 0 effective rank grows from 6.86 (q = 4) to 13.94 (q = 6) to 43.04 (q = 11), tracking the seed 0 F1 improvement from 0.488 to 0.504 to 0.586. Each additional qubit adds informative kernel directions that the SVM can exploit. Beyond q = 11, kernel concentration begins to dominate on seed 0: the eigenvalues flatten toward uniformity, the effective rank saturates, and the SVM loses discriminative
11 signal. F1 collapses at q = 16 on seed 0; multi-seed mean is 0.377, a Tier-1 win. This rise-peak-collapse pattern, mediated by effective rank, constitutes an empirically grounded explanation for the quantum advantage window.
D.
formance peak (q=11) reflects clinically meaningful variation or amplified demographic confounding. Tools such as projected kernels, attention-based localization, and counterfactual auditing [12] offer candidate methodologies for this analysis.
Normalization as a Design Principle F.
Trace normalization plays a role analogous to batch normalization in deep learning: it ensures that the kernel matrix is well-conditioned before being passed to the SVM solver. Frobenius normalization divides by the global Frobenius norm of the kernel matrix, which is dominated by the large diagonal entries and effectively suppresses all off-diagonal information and collapses the kernel to a near-identity structure. Practitioners building quantum kernel pipelines should treat normalization as a primary hyperparameter.
E.
Latent Socioeconomic Signal and Implications for Fairness
The insurance classification task studied here sits at the intersection of two distinct concerns. Methodologically, insurance status serves as a proxy for a subtle, distributed signal that stress-tests kernel expressiveness under class imbalance. A separate and more unsettling question also arises. The fact that this signal is recoverable from chest radiographs at all, by both classical and quantum models, implies that medical images encode socioeconomic stratification in ways that neither clinicians nor patients are aware of. This encoding likely reflects spurious correlations rather than direct causal pathways: differences in acquisition equipment across hospital systems, site-specific positioning conventions, and cumulative markers of environmental or occupational exposure that covary with insurance type without bearing any direct relationship to the underlying pathology [10, 11, 14]. The same foundation models used here (RAD-DINO, MedSigLIP) have been shown to encode demographic attributes in a companion study on shortcut learning in medical imaging [30]. If the discriminative signal is demographic rather than clinical in nature, any sufficiently expressive model trained on such data risks learning these latent signals. Classifier errors then concentrate disproportionately on underrepresented groups even when aggregate performance appears adequate [13, 14]. The quantum advantage demonstrated here sharpens this concern rather than resolving it. A kernel method with higher effective rank (the property that allows QSVM to avoid majority-class collapse) is also better positioned to exploit subtle spurious structure. Interpretability and auditing should therefore be first-class requirements when deploying quantum kernel methods in clinical settings. Future work should examine what structure the quantum feature map is exploiting: whether the discriminative signal captured by the quantum feature map at the per-
Limitations
Spurious signal. Predicting insurance type from chest radiographs may rely in part on spurious correlations: acquisition artifacts, institutional patterns, or demographic proxies encoded in the embeddings. The observed QSVM advantage is therefore best interpreted as improved separability within the representation space, not as evidence of clinically causal signal. Higher-capacity kernels, including the quantum kernel used here, may be better able to exploit this latent structure. Evaluating stability under distribution shift is an important direction for future work. Simulated quantum hardware. All QSVM experiments use Qiskit’s Statevector simulator (exact, noiseless simulation on CPU/GPU). Results on real quantum hardware may differ due to gate errors, decoherence, limited qubit connectivity, and readout noise. Hardware noise exacerbates kernel concentration [18], so the appropriate interpretation is that the BSP circuit architecture has the capacity for advantage in noiseless simulation, not that advantage has been demonstrated on a physical quantum computer. Single dataset and single center. All results derive from the insurance classification task on MIMIC-CXR, collected at Beth Israel Deaconess Medical Center in Boston. The 70/30 Medicare-Medicaid versus Private payer mix reflects Massachusetts, a near-universal-coverage setting, and may not generalise to institutions with different payer structures or outside the United States. Extending to other prediction tasks, patient populations, and healthcare systems will require additional validation. SVM-only classical baselines. The classical collapse documented here is specific to kernel SVMs operating on low-rank PCA representations. Non-kernel classifiers (gradient-boosted trees, logistic regression, or shallow neural networks) may not exhibit the same failure mode and could set a stronger classical ceiling. The quantum kernel advantage is therefore relative to SVM-based baselines; extending the comparison to non-kernel methods is an important open question. Task scope. A quantum advantage at predicting insurance status, a demographic proxy, is a narrower claim than advantage on a clinically meaningful diagnostic task. Until the discriminative signal is shown to reflect genuine clinical variation rather than acquisition artifacts or demographic confounding (§ VI), the result should be interpreted as a methodological finding about kernel expressiveness, not as evidence of clinical utility. Preprocessing stratum selection. DT9 was selected as the primary preprocessing stratum because it produced
12 the strongest quantum results in preliminary experiments across multiple strata. The non-collapse Tier-1 advantage is therefore DT9-specific until multi-strata validation confirms it generalises to other preprocessing configurations. Rank-matched classical kernel. The QSVM advantage may reflect the specific spectral structure of the quantum kernel rather than its effective rank alone (§ VI A). To isolate these factors, we ran a rank-matched RBF experiment on MedSigLIP-448 at q ∈ {4, 6, 11, 16} across all 10 embedding seeds: for each seed, γ ∗ is chosen by binary search so that eff rank(RBF(γ ∗ )) = eff rank(KQ ), using the seed-0 quantum kernel as the fixed effective-rank target. Results are in Table X. At q = 4, rank-matched RBF collapses on 3/10 seeds (more than QSVM’s 2/10) despite having the same effective rank. Mean F1 is 0.110 for rank-matched RBF vs. 0.212 for QSVM. At q = 6, all three methods collapse on 2/10 seeds, but QSVM mean F1 (0.286) is still 68% higher than rank-matched RBF (0.171). At q = 11 and q = 16, neither RBF variant collapses; QSVM achieves mean F1 of 0.343 and 0.377 vs. 0.304 and 0.321 for rank-matched RBF. Two conclusions follow. First, matching the quantum kernel’s effective rank does not reproduce its collapseavoidance: rank-matched RBF is more collapse-prone than QSVM at q = 4 despite identical effective rank. Second, QSVM outperforms rank-matched RBF on mean F1 at every qubit count, by margins of 0.056–0.115. Together these results indicate that the QSVM advantage is not attributable to effective rank alone; the specific spectral structure of the quantum feature map (eigenvalue distribution and off-diagonal correlations not captured by the Shannon effective rank) contributes to both collapse resistance and predictive performance. Setting γ ∗ also requires prior knowledge of KQ , unavailable at deployment; QSVM achieves superior geometry without per-seed tuning. TABLE X. Rank-matched RBF vs. QSVM across 10 seeds (MedSigLIP-448). Collapse = fraction of seeds with F1 < 0.05. γ ∗ is binary-searched per seed to satisfy eff rank(RBF(γ ∗ )) = eff rank(KQ ). All methods use C = 1. QSVM F1 from multiseed runs. eff rank(KQ ) values are the seed-0 fixed targets used for γ ∗ binary search; they differ from Table V (e.g. 43.04 vs. 69.80 at q = 11) because the two experiments use different pre-computed kernel matrices (distinct data-preparation pipelines).
sification on MIMIC-CXR chest radiographs under noiseless simulation. Under a rigorous two-tier fair comparison framework, QSVM wins all 18 Tier-1 configurations on minority-class F1 (17 at p < 0.001, 1 at p < 0.01; paired bootstrap) and all seven Tier-2 F1 comparisons (7/7). Most wins occur in configurations where the classical linear SVM collapses to F1 = 0 on 90–100 % of seeds, rather than in direct competition with a functional baseline. The core mechanism is structural: the linear classical kernel KL has effective rank equal to the PCA dimension q (3.77–5.85 out of 1,896 training samples), which causes irreversible majority-class collapse that is invariant to the regularization parameter C. Multi-seed analysis reveals this collapse is pervasive: classical linear SVM collapses to F1 = 0 on 90–100% of seeds at every qubit count tested across all three models. Direct measurement of the 1,896×1,896 quantum kernel matrix (seed 0) confirms that the quantum feature map achieves Shannon effective ranks of 6.86 and 13.94 at q = 4 and q = 6 (1.82× and 2.52× the linear values), with the ratio growing with qubit count as the quantum feature map accesses an exponentially larger Hilbert space. Beyond the primary finding, our ablation studies yield three practical design recommendations for quantum kernel practitioners: (1) trace normalization is necessary for meaningful F1 and should be treated as a primary pipeline hyperparameter; (2) the qubit count–performance curve can be non-monotonic on individual seeds: on seed 0, MedSigLIP-448 drops sharply at q = 16 (F1 0.586→0.173), while multi-seed mean rises to 0.377 at q = 16 (a Tier1 win); practitioners should validate on multiple seeds before inferring a performance peak. This suggests that barren plateau effects [18, 21] emerge before hardware limits are reached in some seeds; (3) 1-DOF angle encoding outperforms 3-DOF, and deeper re-uploading (reps=2) degrades performance at q = 8; circuit expressivity and sample size must be co-designed. As quantum hardware matures and simulation capacity grows to larger qubit counts, the quantum advantage demonstrated here on a real-world medical imaging task at q ≤ 16 provides a foundation for future work on noiseaware quantum kernels, and extension to other medical imaging modalities. The concentration observed at q = 16 suggests that scaling QSVM beyond this regime will require more than adding qubits. Supplementary figures covering all models and qubit counts are provided in Appendix A.
q eff rank(KQ ) Collapse (RBFscale ) Collapse (RBF∗ ) Collapse (QSVM) F1 (RBFscale ) F1 (RBF∗ ) F1 (QSVM) 4 6.94 0.2 0.3 0.2 0.169 0.110 0.212 6 14.09 0.2 0.2 0.2 0.178 0.171 0.286 11 69.80 0.0 0.0 0.1 0.272 0.304 0.343 16 250.99 0.0 0.0 0.1 0.269 0.321 0.377
ACKNOWLEDGMENTS
VII.
CONCLUSION
Across 10 embedding seeds and three medical foundation models, QSVM with frozen embeddings provides evidence of quantum advantage in binary insurance clas-
This work was supported by the Google Cloud Research Credits program under award number GCP19980904. The MIMIC-CXR-JPG dataset used in this study was obtained from PhysioNet under a data use agreement (credentialed access). D.E.K. is supported by the Agency for Science, Technology and Research (A*STAR) under the Quantum Innovation Centre (Q.InC) Strategic Research and
13 Translational Thrust (SRTT). S.T.G. acknowledges the support from the National Research Foundation, Singapore through the National Quantum Office, hosted in Agency for Science, Technology and Research (A*STAR), Singapore under its Quantum Engineering Programme 3.0 Funding Initiative (W24Q3D0002). The authors thank the MIT Critical Data community for support and discussion. Computational resources were provided by the MIT Office of Research Computing and Data (ORCD) through A100 and H200 GPU allocations. L.A.C. is funded by the National Institutes of Health through NIBIB R01 EB017205. RG is supported by the Johns Hopkins Institute for Clinical and Translational Research (ICTR) and the National Center for Advancing Translational Sciences (NCATS), National Institutes of Health (NIH) grant number T32TR004928. The contents of this publication are solely the responsibility of the authors and do not necessarily represent the official view of the Johns Hopkins ICTR, NCATS or NIH.
REFERENCES [1] Vojtěch Havlı́ček, Antonio D. Córcoles, Kristan Temme, Aram W. Harrow, Abhinav Kandala, Jerry M. Chow, and Jay M. Gambetta. Supervised learning with quantumenhanced feature spaces. Nature, 567(7747):209–212, 2019. [2] M. Schuld and N. Killoran. Quantum machine learning in feature Hilbert spaces. Physical Review Letters, 122:040504, 2019. [3] M. Schuld. Supervised quantum machine learning models are kernel methods, 2021. [4] Yunchao Liu, Srinivasan Arunachalam, and Kristan Temme. A rigorous and robust quantum speed-up in supervised machine learning. Nature Physics, 17(9):1013– 1017, 2021. [5] S. Jerbi, L. J. Fiderer, H. Poulsen Nautrup, J. M. Kübler, H. J. Briegel, and V. Dunjko. Quantum machine learning beyond kernel methods. Nature Communications, 14:517, 2023. [6] Joseph Bowles, Shahnawaz Ahmed, and Maria Schuld. Better than classical? the subtle art of benchmarking quantum machine learning models. arXiv preprint arXiv:2403.07059, 2024. [7] Sebastián Andrés Cajas Ordóñez, Luis Fernando Torres Torres, Mario Bifulco, Carlos Andres Duran, Cristian Bosch, and Ricardo Simon Carbajo. Embedding aware quantum classical svms for scalable quantum machine learning. In Marco Baioletti, Miguel Angel Gonzalez, Corrado Loglisci, Angelo Oddi, Riccardo Rasconi, and Ramiro Varela, editors, Proceedings of the 3rd International Workshop on AI for Quantum and Quantum for AI (AIQxQIA 2025), co-located with the 28th European Conference on Artificial Intelligence (ECAI 2025), volume 4153 of CEUR Workshop Proceedings, Bologna, Italy, October 2025. CEUR-WS.org. [8] Alistair E. W. Johnson, Tom J. Pollard, Seth J. Berkowitz, Nathaniel R. Greenbaum, Matthew P. Lungren, ChihYing Deng, Roger G. Mark, and Steven Horng. MIMICCXR, a de-identified publicly available database of chest
radiographs with free-text reports. Scientific Data, 6:317, 2019. [9] A. E. W. Johnson, T. J. Pollard, N. R. Greenbaum, M. P. Lungren, C.-Y. Deng, Y. Peng, Z. Lu, R. G. Mark, S. J. Berkowitz, and S. Horng. MIMIC-CXR-JPG, a large publicly available database of labeled chest radiographs, 2019. [10] Judy Wawira Gichoya, Imon Banerjee, Ananth Reddy Bhimireddy, John L. Burns, Leo Anthony Celi, Li-Ching Chen, Ramon Correa, Natalie Dullerud, Marzyeh Ghassemi, Shih-Cheng Huang, Po-Chih Kuo, Matthew P. Lungren, Lyle J. Palmer, Brandon J. Price, Saptarshi Purkayastha, Ayis T. Pyrros, Lauren Oakden-Rayner, Chima Okechukwu, Laleh Seyyed-Kalantari, Hari Trivedi, Ryan Wang, Zachary Zaiman, and Haoran Zhang. AI recognition of patient race in medical imaging: a modelling study. The Lancet Digital Health, 4(6):e406–e414, 2022. [11] Chi-Yu Chen, Rawan Abulibdeh, Arash Asgari, Sebastián Andrés Cajas Ordóñez, Leo Anthony Celi, Deirdre Goode, Hassan Hamidi, Laleh Seyyed-Kalantari, Ned McCague, Thomas Sounack, et al. Algorithms trained on normal chest x-rays can predict health insurance types. arXiv preprint arXiv:2511.11030, 2025. [12] Charles Jones, Daniel C Castro, Fabio De Sousa Ribeiro, Ozan Oktay, Melissa McCradden, and Ben Glocker. A causal perspective on dataset bias in machine learning for medical imaging. Nature Machine Intelligence, 6(2):138– 146, 2024. [13] Laleh Seyyed-Kalantari, Haoran Zhang, Matthew B. A. McDermott, Irene Y. Chen, and Marzyeh Ghassemi. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nature Medicine, 27(12):2176–2182, 2021. [14] Ziad Obermeyer, Brian Powers, Christine Vogeli, and Sendhil Mullainathan. Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464):447–453, 2019. [15] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975–11986, 2023. [16] Fernando Pérez-Garcı́a, Harshita Sharma, Sam BondTaylor, Kenza Bouzid, Valentina Salvatelli, Maximilian Ilse, Shruthi Bannur, Daniel C Castro, Anton Schwaighofer, Matthew P Lungren, et al. Exploring scalable medical image encoders beyond text supervision. Nature Machine Intelligence, 7(1):119–130, 2025. [17] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby. An image is worth 16×16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. [18] Supanut Thanasilp, Samson Wang, Marco Cerezo, and Zoë Holmes. Exponential concentration in quantum kernel methods. Nature Communications, 15(1):5200, 2024. [19] H.-Y. Huang, M. Broughton, M. Mohseni, R. Babbush, S. Boixo, H. Neven, and J. R. McClean. Power of data in quantum machine learning. Nature Communications, 12:2631, 2021. [20] Jonas Kübler, Simon Buchholz, and Bernhard Schölkopf. The inductive bias of quantum kernels. Advances in Neural Information Processing Systems, 34:12661–12673,
14 2021. [21] M. Larocca, S. Thanasilp, S. Wang, K. Sharma, J. Biamonte, P. J. Coles, L. Cincio, J. R. McClean, Z. Holmes, and M. Cerezo. Barren plateaus in variational quantum computing. Nature Reviews Physics, 7:174–189, 2025. [22] Amira Abbas, David Sutter, Christa Zoufal, Aurelien Lucchi, Alessio Figalli, and Stefan Woerner. The power of quantum neural networks. Nature Computational Science, 1:403–409, 2021. [23] D. Peral-Garcı́a, J. Cruz-Benito, and F. J. Garcı́a-Peñalvo. Systematic literature review: Quantum machine learning and its applications. Computer Science Review, 51:100619, 2024. [24] A. Senokosov, A. Sedykh, A. Sagingalieva, B. Kyriacou, and A. Melnikov. Quantum machine learning for image classification. Machine Learning: Science and Technology, 5:015040, 2024. [25] Vladimir N. Vapnik. The Nature of Statistical Learning Theory. Springer, New York, 2nd edition, 1998. [26] B. Schölkopf and A. J. Smola. Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond. MIT Press, Cambridge, MA, 2002. [27] C. Coleman, C. Yeh, S. Mussmann, B. Mirzasoleiman, P. Bailis, P. Liang, J. Leskovec, and M. Zaharia. Selection via proxy: Efficient data selection for deep learning. In Proceedings of the International Conference on Learning Representations (ICLR), 2020. [28] M. Sokolova and G. Lapalme. A systematic analysis of performance measures for classification tasks. Information Processing & Management, 45(4):427–437, 2009. [29] Davide Chicco and Giuseppe Jurman. The advantages of the Matthews correlation coefficient (MCC) over F1 score and accuracy in binary classification evaluation. BMC Genomics, 21:6, 2020. [30] Felipe Ocampo Osorio, Santiago Pedroza Gomez, David Esteban Rebellón Sanchez, Richard Ramirez Fernandez, Reinel Tabares-Soto, Mario Alejandro Bravo-Ortı́z, and Gustavo Adolfo Cruz Suarez. Predicting no-shows at outpatient appointments in internal medicine using machine learning models. PeerJ Computer Science, 11:e2762, 2025. [31] Jarrod R McClean, Sergio Boixo, Vadim N Smelyanskiy, Ryan Babbush, and Hartmut Neven. Barren plateaus in quantum neural network training landscapes. Nature Communications, 9(1):4812, 2018.
2.
Quantum Kernel Heatmaps (All Models)
Figure 9 shows the quantum kernel matrices KQ at q = 4 and q = 6 for all three models.
3.
PCA Feature Space: Class Separation at q = 4 and q = 6
Figure 10 shows the PCA-compressed training data at q = 4 and q = 6 for all three models. The substantial class overlap visible in every panel provides a geometric explanation for why the linear kernel collapses.
4.
PCA Geometry of MedSigLIP-448 at q = 2 5.
ViT-patch32-GAP Pooling Ablation
To assess the effect of pooling strategy on quantum kernel performance, we evaluate a global average pooling (GAP) variant of ViT-patch32 alongside the CLS-token variant reported in the main text. Both variants produce 768-dimensional embeddings from the same frozen ViTpatch32 backbone; the only difference is the aggregation of patch tokens: GAP averages all patch tokens, while CLS uses only the class token. Multi-seed experiments (10 seeds, DT9, trace normalization, C=1) were completed for both variants across 11 qubit configurations (q ∈ {2, 3, 4, 5, 6, 8, 9, 10, 11, 12, 16}). Table XI reports QSVM minority-class F1 (mean ± std over 10 seeds) for both pooling variants, together with the best classical SVM baseline (RBF kernel, C=1) evaluated on the same splits. The two pooling strategies yield nearly identical QSVM performance at q ≥ 10 (difference ≤ 0.003), with CLS slightly higher on average across all q. Both variants show a quantum advantage over the best classical baseline for q ≥ 4 under noiseless simulation. The CLS variant was selected for the main text because it matches the standard ViT evaluation protocol and shows marginally more consistent F1 across the full qubit sweep.
Appendix A: Supplementary Figures 6.
This appendix collects additional figures that complement the main text. All experiments use DT9 preprocessing, seed 0, and trace normalization unless noted otherwise.
1.
Quantum Kernel Eigenspectra (All Models)
Figure 6 shows the quantum kernel eigenvalue spectra for all three embedding models at q = 4 and q = 6. These complement the MedSigLIP q = 6 spectrum shown in the main text (Figure 3).
ViT-patch16-cls Patch-Size Ablation
To assess the effect of patch size on quantum kernel performance, we evaluate a ViT with patch size 16 (ViT-patch16-cls, 768-dimensional CLS-token embeddings) alongside the ViT-patch32-cls variant reported in the main text. Both variants use the same frozen ViT backbone architecture; the only difference is the spatial resolution of the patch tokenisation (patch16 produces 4× more tokens per image than patch32 and yields a denser spatial representation before CLS pooling). Multi-seed experiments (10 seeds, DT9, trace normalization, C=1) were completed for ViT-patch16-cls across the same 11 qubit configurations.
15
FIG. 6. Quantum kernel eigenvalue spectra for all three embedding models at q = 4 (left) and q = 6 (right). MedSigLIP-448 (top), RAD-DINO (middle), ViT-patch32-cls (bottom). The quantum kernel consistently exhibits higher effective rank than the linear kernel across all models and qubit counts.
FIG. 7. Quantum kernel eigenspectrum for MedSigLIP-448 at the performance peak q = 11 (seed 0). Shannon effective rank = 43.04, a 6.3× increase from q = 4 (6.86); multi-seed mean is 69.80.
FIG. 8. Quantum kernel heatmap for MedSigLIP-448 at q = 11, seed 0 (200 training samples). Compare with q = 6 (Figure 9): the q = 11 kernel shows richer off-diagonal structure, consistent with its higher effective rank (43.04 vs. 13.94, seed 0).
Table XII shows that ViT-patch16-cls yields substan-
tially lower QSVM minority-class F1 than ViT-patch32-
16
FIG. 9. Quantum kernel heatmaps for all three embedding models at q = 4 (left) and q = 6 (right). The block structure reflects class boundaries in the training data (sorted by label). Higher qubit counts show sharper off-diagonal structure, consistent with increased effective rank.
cls at every qubit count tested (∆F1 ≈ −0.24 at q = 16,
−0.28 at q = 8). This performance gap is the primary
17 TABLE XI. CLS vs. GAP pooling: QSVM minority-class F1 (mean ± std, 10 seeds) and best classical SVM baseline (RBF, C=1) for ViT-patch32 on DT9. Both pooling variants produce 768-dimensional embeddings. ∆GAP = QSVM-GAP − BestClassical. q 2 3 4 5 6 8 9 10 11 12 16
QSVM-CLS 0.016 ± 0.028 0.003 ± 0.008 0.048 ± 0.073 0.212 ± 0.101 0.272 ± 0.132 0.370 ± 0.092 0.384 ± 0.081 0.391 ± 0.081 0.403 ± 0.064 0.405 ± 0.063 0.399 ± 0.098
QSVM-GAP Best-Cl ∆CLS ∆GAP 0.000 ± 0.000 0.033 −0.018 −0.033 0.008 ± 0.025 0.041 −0.038 −0.018 0.102 ± 0.092 0.044 +0.004 +0.058 0.205 ± 0.110 0.180 +0.032 +0.025 0.235 ± 0.090 0.219 +0.054 +0.016 0.278 ± 0.122 0.251 +0.119 +0.027 0.276 ± 0.141 0.272 +0.112 +0.004 0.383 ± 0.069 0.287 +0.103 +0.096 0.406 ± 0.092 0.303 +0.100 +0.103 0.412 ± 0.082 0.307 +0.098 +0.105 0.397 ± 0.082 0.330 +0.104 +0.067
basis for selecting ViT-patch32-cls as the main-text ViT baseline. The result suggests that the denser patch16 representation introduces additional redundancy or noise
in the low-dimensional PCA subspace, making it harder for the quantum kernel to separate insurance classes. TABLE XII. Patch-size ablation: QSVM minority-class F1 (mean ± std, 10 seeds) for ViT-patch32-cls and ViT-patch16cls on DT9. Both use CLS-token embeddings; best classical SVM baseline (RBF, C=1) shown for ViT-patch32-cls (the main-text model). q QSVM patch32 QSVM patch16 ∆ (patch32 − patch16) 2 0.016 ± 0.028 0.000 ± 0.000 +0.016 +0.003 3 0.003 ± 0.008 0.000 ± 0.000 4 0.048 ± 0.073 0.008 ± 0.024 +0.040 5 0.212 ± 0.101 0.005 ± 0.016 +0.207 +0.251 6 0.272 ± 0.132 0.021 ± 0.031 8 0.370 ± 0.092 0.069 ± 0.110 +0.301 +0.305 9 0.384 ± 0.081 0.079 ± 0.118 10 0.391 ± 0.081 0.080 ± 0.148 +0.311 11 0.403 ± 0.064 0.121 ± 0.147 +0.282 12 0.405 ± 0.063 0.104 ± 0.159 +0.301 16 0.399 ± 0.098 0.160 ± 0.118 +0.239
18
FIG. 10. PCA feature space (q = 4, left; q = 6, right) for all three models (MedSigLIP-448, RAD-DINO, ViTpatch32-cls, top to bottom). Each point is a training sample projected onto its first two PCA components after StandardScaler→PCA→MinMaxScaler preprocessing (seed 0). Blue: Medicaid/Medicare; orange: Private insurance. The two classes overlap substantially in every panel, which explains why the linear kernel KL —operating in this same q-dimensional subspace—collapses to majority-class prediction (F1 = 0). The label in the lower-left corner of each panel confirms classical SVM collapse at that configuration.
19
FIG. 11. PCA scatter of MedSigLIP-448 embeddings projected to 2 components (total explained variance: 21.8%). Train set: 1319 majority (Medicare) vs 577 minority (Private). The low explained variance confirms that the 2D PCA projection captures only a fraction of the structure exploited by the quantum kernel in higher dimensions.