Skip to content

Quantum Kernel Advantage over Classical Collapse in Medical Foundation Model Embeddings

🕒 Published (v1): 2026-04-27 00:00 UTC · Source: HuggingFace · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This paper demonstrates quantum kernel advantage over classical SVMs on a real clinical imaging task — binary insurance classification from MIMIC-CXR chest radiographs — using frozen embeddings from three medical foundation models (MedSigLIP-448, RAD-DINO, ViT-patch32). QSVM wins minority-class F1 in all 18 tested configurations versus an untuned linear SVM, and in all 7 configurations versus a C-tuned RBF SVM. The advantage is mechanistically explained by the quantum kernel's higher effective rank in the same PCA-compressed subspace where the classical kernel structurally collapses.

Problem

Classical kernel SVMs applied to PCA-compressed foundation model embeddings collapse to majority-class prediction due to structural rank deficiency: after PCA reduction to q dimensions, the linear kernel matrix has effective rank ≈ q out of N ≈ 1,896 training samples, making minority-class separation impossible regardless of regularization. Existing QML studies lack rigorous fair comparisons and operate on synthetic or toy datasets rather than real clinical imaging data with class imbalance.

Method

Frozen CLS-token embeddings from MedSigLIP-448 (448-d), RAD-DINO (768-d), and ViT-patch32 (768-d) are compressed via StandardScaler → PCA-q → MinMaxScaler[−1,1] to q ∈ {4,6,8,9,10,11,12,16} dimensions. A Block-Sparse Parameterization (BSP) circuit with 1-DOF angle encoding (single Ry per qubit) and ring CNOT entanglement computes the quantum kernel via the compute-uncompute strategy: K_Q(u_i, u_j) = |⟨0^q|U†(u_i)U(u_j)|0^q⟩|². Trace normalization (K̃_Q = K_Q / tr(K_Q)) is applied before passing to the SVM solver. A two-tier fair comparison framework is used: Tier 1 (untuned QSVM C=1 vs. untuned linear SVM C=1) and Tier 2 (untuned QSVM vs. C-tuned RBF SVM). Statistical validation uses 10 embedding seeds with paired bootstrap (10,000 resamples). Kernel effective rank (Shannon entropy of normalized eigenspectrum) is computed to explain collapse.

Key Contributions

  • 18/18 Tier-1 wins: QSVM beats untuned linear SVM on mean minority-class F1 across all model×qubit configurations, validated over 10 seeds (17 at p < 0.001, 1 at p < 0.01).
  • 7/7 Tier-2 wins: Untuned QSVM beats C-tuned RBF SVM at identical PCA dimensionality (mean gain +0.068 F1, max +0.112).
  • Structural collapse explanation: Linear kernel effective rank equals exactly q after PCA-q compression; quantum kernel effective rank reaches 43.04–69.80 at q = 11, a 4–6× ratio, explaining why QSVM maintains non-trivial minority-class recall where classical kernels fail.
  • Three circuit design rules: (1) Trace normalization is necessary (Frobenius normalization collapses F1 to 0); (2) 1-DOF Ry encoding outperforms 3-DOF Rz-Ry-Rz; (3) increasing re-uploading depth (reps > 1) degrades performance — sample size, not circuit capacity, is the bottleneck.
  • Architecture-dependent concentration: A full qubit sweep reveals model-specific barren-plateau behavior — MedSigLIP-448 peaks at q = 11 then collapses at q = 16 (seed-dependent), while RAD-DINO and ViT-patch32 improve monotonically.

Results

  • Strongest configuration: MedSigLIP-448 at q = 11 — QSVM mean F1 = 0.343 ± 0.170 vs. linear SVM F1 = 0.050 ± 0.159 (ΔF1 = +0.293, 95% CI [+0.190, +0.385], p < 0.001).
  • Classical linear SVM collapses to F1 = 0 on 90–100% of seeds at every qubit count across all models; C-invariant over C ∈ {0.01, 0.1, 1, 10, 100}.
  • Tier-2 (vs. C-tuned RBF): QSVM wins all 7 configurations; gains range from +0.004 (ViT-p32 q=4) to +0.112 (MedSigLIP q=8), relative gains +9%–+61%.
  • RAD-DINO QSVM achieves statistically significant accuracy gains at q ∈ {4,6,8,10} (Δacc = +1.2–1.8%, p ≤ 0.005).
  • Rank-matched RBF experiment (γ tuned to match QSVM effective rank): mean F1 = 0.110 vs. QSVM 0.212 at q=4, with 30% RBF collapse vs. 20% QSVM collapse — confirms quantum spectral structure contributes beyond rank alone.
  • Projected quantum kernel at q = 16 recovers MedSigLIP F1 from 0.173 to 0.396, confirming fidelity measurement concentration (not circuit expressibility) is the bottleneck at high qubit counts.

Limitations

  • Results are validated only on the DT9 preprocessing stratum of MIMIC-CXR, which was selected post-hoc based on strong preliminary quantum results; generalization to other strata unconfirmed.
  • Dataset is limited to N = 2,371 samples; quantum hardware constraints cap practical training at ~2,000 samples with q ≤ 16 qubits.
  • All experiments use noiseless statevector simulation; real quantum hardware noise is not modeled.
  • Insurance classification is a socioeconomic proxy task, not a clinical prediction target; the paper explicitly notes it is not intended for clinical deployment.
  • Tier-2 comparison tunes C but not RBF bandwidth γ at fixed PCA-q, leaving open the question of whether full classical hyperparameter search (including γ) would close the gap.
  • The q = 16 projected kernel recovery result is mechanistic/diagnostic and tested only at seed 0; cross-seed and cross-model generalization is not established.
  • Single circuit architecture (BSP 1-DOF) evaluated; results may not transfer to other quantum circuit families.

Relevance to Foundation Models in Medicine

This paper is the first to use frozen embeddings from domain-specific medical foundation models (MedSigLIP-448, RAD-DINO) as the interface between quantum kernel methods and clinical imaging, revealing that the embedding geometry itself determines whether quantum advantage manifests — architecture-dependent concentration rates differ across models. For researchers tracking foundation models in medicine, it establishes that the representational structure of medical FMs creates a regime (low PCA-q) in which classical kernel methods fundamentally fail on imbalanced tasks, and quantum kernels provide a practical remedy without fine-tuning. The work also surfaces a health-equity dimension: socioeconomic signals encoded in medical FM embeddings are recoverable by quantum methods even when classical methods cannot separate the minority class, raising questions about what information medical FMs capture and propagate.