Quantum Kernel Advantage over Classical Collapse in Medical Foundation Model Embeddings

1MIT Critical Data, Massachusetts Institute of Technology, Cambridge, MA, USA   2Clinical Research Center, Artificial Intelligence Unit, Fundación Valle del Lili, Cali, Valle del Cauca, Colombia   3Quantum Innovation Centre (Q.InC), Agency for Science, Technology and Research (A*STAR), Singapore 138634   4Institute of High Performance Computing (IHPC), Agency for Science, Technology and Research (A*STAR), Singapore 138632
5Science, Mathematics and Technology Cluster, Singapore University of Technology and Design, Singapore 487372   6Department of Electronics, Information and Bioengineering, Politecnico di Milano, Milan, Italy   7Bordeaux Population Health Research Center, Inserm U1219, Université de Bordeaux, F-33000 Bordeaux, France   8Inria Bordeaux, Université de Bordeaux, F-33000 Bordeaux, France
9Universidad del Cauca, Popayán, Colombia   10Singapore Management University, 81 Victoria St, Singapore 188065   11National Taiwan University Hospital   12School of Medicine, Johns Hopkins University, Baltimore, MD, USA
13Department of Biomedical Engineering, Johns Hopkins University, Baltimore, MD, USA   14AI for Responsible, Generalizable, and Open Surgical (ARGOS) Research Group, Baltimore, MD, USA   15School of Information Engineering, Shanghai Maritime University, Shanghai 201306, China
16Research Center of Intelligent Information Processing and Quantum Intelligent Computing, Shanghai 201306, China   17Laboratory for Computational Physiology, MIT, Cambridge, MA, USA   18Department of Medicine, Beth Israel Deaconess Medical Center, Boston, MA, USA

*Corresponding Author

🎧 Paper Summary  —  listen to a 5-minute audio overview

Key Results

  • 18/18 Tier-1 wins: QSVM beats untuned linear SVM on minority-class F1 across all configurations (17 at p<0.001, 1 at p<0.01)
  • At q=11, MedSigLIP-448: mean F1 = 0.343 ± 0.170 vs classical F1 = 0.050 ± 0.159 (ΔF1 = +0.293, p<0.001)
  • 7/7 Tier-2 wins: QSVM beats C-tuned RBF SVM (mean gain +0.068)
  • Quantum kernel effective rank reaches 69.80 at q=11, versus linear kernel rank of exactly 11
  • Classical collapse is C-invariant: persists regardless of regularization strength

Abstract

We investigate whether quantum kernel methods can outperform classical kernels on binary insurance classification using MIMIC-CXR chest radiographs, where class imbalance causes standard linear classifiers to collapse. We evaluate quantum support vector machines (QSVM) with frozen embeddings from three medical foundation models (MedSigLIP-448, RAD-DINO, ViT-patch32) under noiseless simulation, using a two-tier fair comparison framework.

At Tier 1 (untuned QSVM vs. untuned linear SVM, C=1), QSVM wins minority-class F1 in all 18 configurations (17 at p<0.001, 1 at p<0.01). The classical linear kernel collapses to F1=0 on 90–100% of seeds at every qubit count, while QSVM maintains recall. At q=11, QSVM achieves mean F1=0.343±0.170 versus classical F1=0.050±0.159 (ΔF1=+0.293, p<0.001). At Tier 2 (untuned QSVM vs. C-tuned RBF SVM), QSVM wins all 7 configurations (mean gain +0.068).

Eigenspectrum analysis shows quantum kernel effective rank reaches 69.80 at q=11, far exceeding linear kernel rank of exactly q=11. These results provide evidence of quantum kernel advantage on a real medical imaging task. The identified mechanism — kernel rank collapse in classical methods under class imbalance — gives a concrete basis for predicting where quantum kernels hold an edge.

Results

F1 score vs qubit count across medical foundation models

F1 score vs. qubit count across models. QSVM minority-class F1 scales with qubit count across MedSigLIP-448, RAD-DINO, and ViT-patch32 embeddings, while the classical linear SVM remains collapsed near zero.

Summary bar chart comparing QSVM and classical SVM across all configurations

Summary comparison across all 18 Tier-1 configurations. QSVM (green) consistently outperforms the classical linear SVM (gray) on minority-class F1. The classical kernel collapses to F1=0 on 90–100% of seeds at every qubit count.

Quantum kernel eigenspectrum at q=11 for MedSigLIP-448

Quantum kernel eigenspectrum at q=11 (MedSigLIP-448). The quantum kernel achieves an effective rank of 69.80, spreading eigenvalue mass across many dimensions. The linear kernel's rank equals exactly q=11 — insufficient to resolve the minority class under severe imbalance.

Quantum vs linear eigenspectrum comparison at q=4 and q=6

Quantum vs. linear eigenspectrum at q=4 and q=6. Across qubit counts, the quantum kernel maintains a richer eigenspectrum than the linear kernel, which concentrates its rank budget and fails on imbalanced data.

Quantum kernel matrix heatmap at q=11 for MedSigLIP-448

Quantum kernel matrix, q=11 (MedSigLIP-448). Rich off-diagonal block structure reflects class boundaries preserved by the quantum feature map. Effective rank = 43.04 (multi-seed mean 69.80).

Quantum kernel matrix heatmap at q=16 for MedSigLIP-448

Quantum kernel matrix, q=16 (MedSigLIP-448). At q=16, swap-test fidelity concentration causes off-diagonal entries to converge toward a single value, consistent with the onset of concentration collapse.

Quantum kernel heatmap RAD-DINO q=4

Kernel heatmap, RAD-DINO q=4.

Quantum kernel heatmap RAD-DINO q=6

Kernel heatmap, RAD-DINO q=6.

Quantum kernel heatmap ViT-patch32 q=4

Kernel heatmap, ViT-patch32 q=4.

Quantum kernel heatmap ViT-patch32 q=6

Kernel heatmap, ViT-patch32 q=6.

Quantum kernel eigenspectrum at q=16 for MedSigLIP-448

Quantum kernel eigenspectrum at q=16 (MedSigLIP-448). At q=16, effective rank rises to 92.13 yet multi-seed mean F1 remains 0.377 (a Tier-1 win), confirming that concentration collapse at q=16 is seed-dependent rather than universal.

PCA scatter MedSigLIP-448 q=4

PCA scatter, MedSigLIP-448 q=4.

PCA scatter RAD-DINO q=4

PCA scatter, RAD-DINO q=4.

PCA scatter ViT-patch32 q=4

PCA scatter, ViT-patch32 q=4.

PCA scatter MedSigLIP-448 q=6

PCA scatter, MedSigLIP-448 q=6.

PCA scatter RAD-DINO q=6

PCA scatter, RAD-DINO q=6.

PCA scatter ViT-patch32 q=6

PCA scatter, ViT-patch32 q=6.

Approach

Frozen embeddings from three medical foundation models are extracted from MIMIC-CXR chest radiographs and compressed via PCA to q dimensions, matching the qubit count. A parameterized quantum circuit encodes each compressed embedding as a quantum state; the inner product of pairs of states defines the quantum kernel matrix. A classical SVM then trains on this kernel.

The two-tier comparison framework separates hyperparameter effects from kernel effects. Tier 1 holds the SVM regularization constant (C=1) for both QSVM and linear SVM, isolating the kernel as the only varying factor. Tier 2 allows the classical RBF SVM to tune C via cross-validation, giving it an intentional advantage while keeping QSVM untuned — a conservative test that QSVM still passes in all 7 evaluated cases.

The dataset is a binary insurance classification task drawn from MIMIC-CXR, with severe class imbalance (roughly 9:1). This imbalance is the mechanism that causes classical linear kernels to collapse: their effective rank is bounded by q, the embedding dimension, which is too low to build a separating hyperplane that recovers the minority class.

Citation

@article{cajas2026qml,
  title  = {Quantum Kernel Advantage over Classical Collapse in Medical Foundation Model Embeddings},
  author = {Cajas Ord{\'o}{\~n}ez, Sebasti{\'a}n A. and Ocampo Osorio, Felipe and Koh, Dax Enshan and Al Attrach, Rafi and Marzullo, Aldo and Guerra-Adames, Ariel and Andrade, J. Alejandro and Goh, Siong Thye and Chen, Chi-Yu and Gorijavolu, Rahul and Yang, Xue and Hebdon, Noah Dane and Celi, Leo Anthony},
  journal = {arXiv preprint arXiv:2604.24597},
  year    = {2026},
  url     = {https://arxiv.org/abs/2604.24597}
}