Quantum Medical Patterns

---
name: quantum-medical-patterns
description: "Reusable research patterns from quantum ML in healthcare: hybrid modeling, entanglement PET (J-PET), CV photonic QNNs, quantum autoencoders, quantum ophthalmology, generative models, kernel methods, clinical trials, tensor-network federated learning, TSHF fusion. 4 sub-class skills exist. Triggers: quantum medical, entanglement PET, CV-QNN, quantum autoencoder, quantum ophthalmology, J-PET, 量子医疗."
---

# Quantum Medical Research Patterns

Reusable patterns extracted from arXiv and literature on quantum machine learning
(QML) applications in healthcare, medical imaging, drug discovery, and clinical
decision-making.

## Key Research Patterns

### Pattern 1: Hybrid Quantum-Classical Medical Modeling

**Core idea**: Classical deep learning handles data preprocessing and feature extraction;
quantum circuits process the refined features for classification or regression.

**Typical pipeline**:
```
Raw medical data (EEG/MRI/clinical) → Classical preprocessing → Feature extraction
→ Quantum feature map (angle/ amplitude encoding) → VQC/QNN → Classification
```

**Encoding strategies**:
- **Angle encoding**: Map normalized features to rotation angles of qubits (efficient, low depth)
- **Amplitude encoding**: Embed features into quantum state amplitudes (exponential compression, requires normalization)
- **Basis encoding**: Binary feature representation (simple but qubit-intensive)

**Variational circuits**:
- Use hardware-efficient ansatz for NISQ devices
- 2-4 layers of parameterized rotations + entanglement
- Gradient-based optimization (parameter-shift rule) or gradient-free (SPSA)

**Reference papers**:
- Hybrid Quantum-Classical Model with EEG + Quantum Feature Extraction (Mayo Clinic, 2026)
- Early Detection of Coronary Heart Disease Using Hybrid QML (arxiv)

### Pattern 2: Quantum Generative Models for Medical Imaging

**Core idea**: Quantum Generative Adversarial Networks (QGANs) or quantum variational
autoencoders for medical image enhancement, augmentation, and synthesis.

**Advantages over classical**:
- Lower Fréchet Inception Distance (FID) scores on medical MNIST and knee osteoarthritis X-rays
- Better handling of small medical datasets via quantum expressivity
- Data augmentation for rare disease classes

**Key architectures**:
- **QGAN**: Classical generator + quantum discriminator, or fully quantum both
- **Quantum Circuit Born Machine (QCBM)**: Pure quantum generative model
- **Hybrid VAE**: Classical encoder + quantum latent space + classical decoder

**Reference papers**:
- Quantum Generative Learning for High-Resolution Medical Image Enhancement (IOP, 2025)
- Generative Diffusion Augmentation with Quantum-Enhanced Modeling for Medical Imaging

### Pattern 3: Quantum Kernel Methods for Medical Foundation Models

**Core idea**: Use quantum kernel functions to compute similarity in high-dimensional
Hilbert space, potentially providing advantage over classical kernels for medical data.

**Workflow**:
```
Medical data → Classical embedding (foundation model) → Quantum kernel → QSVM/QSVC
→ Classification/diagnosis
```

**Key finding**: Quantum kernels show advantage when classical kernel methods suffer from
"kernel collapse" — where classical embeddings become indistinguishable in high dimensions.

**Reference papers**:
- Quantum Kernel Advantage over Classical Collapse in Medical Foundation Model Embeddings (arxiv, 2026-04)
- Quantum Machine Learning in Healthcare: Evaluating QNN and QSVM Models (arxiv, 2025)

### Pattern 4: Quantum Optimization for Clinical Trials & Drug Discovery

**Core idea**: Use quantum annealing or QAOA for combinatorial optimization in
patient cohort selection, trial design, and molecular simulation.

**Applications**:
- Patient stratification and cohort matching (QUBO formulation)
- Molecular electronic structure simulation (VQE, adaptive variational algorithms)
- Drug-target interaction prediction (quantum graph neural networks)

**Reference papers**:
- Towards Quantum Computing for Clinical Trial Design (arxiv, 2026-05)
- Convergence Frontier: ML + HPC Quantum Computing for Drug Discovery (arxiv, 2026-05)
- Quantum Simulation of Protein Fragment Electronic Structure (arxiv, 2026)

### Pattern 5: Continuous-Variable Quantum Neural Networks for Biomedical Imaging

**Core idea**: Use continuous-variable (CV) quantum systems (photonic) instead of
discrete qubits for processing high-resolution medical images.

**Advantages**:
- Native compatibility with continuous medical data (pixel intensities, waveforms)
- Higher information density per quantum mode
- Better scalability for image-size inputs

**Reference papers**:
- Towards Continuous-variable Quantum Neural Networks for Biomedical Imaging (arxiv, 2026-05)

### Pattern 6: Temperature-Scaled Hybrid Fusion (TSHF)

**Core idea**: Use a learnable scalar parameter to dynamically balance gradient dynamics
between quantum and classical branches during end-to-end training, resolving optimization
asymmetries that plague naive hybrid architectures.

**Three progressive fusion strategies**:
1. **SHF (Static Hybrid Fusion)**: Offline extraction, simple concatenation. Inflexible.
2. **DHF (Dynamic Hybrid Fusion)**: End-to-end co-adaptation, gradient flows through both branches.
3. **TSHF (Temperature-Scaled Hybrid Fusion)**: Learnable temperature scalars per branch, inspired by multimodal learning. Dynamically balances hybrid gradient dynamics.

**Architecture**:
```
Input → Classical Backbone (ResNet/ViT) → Classical Embedding
    ↘ Quantum Circuit (trainable/deterministic) → Quantum Embedding
        ↓
    TSHF: t_classical, t_quantum (learnable scalars) → Projection → Classifier
```

**Results on BreastMNIST**: TSHF with ResNet + trainable quantum circuit achieved
87.82% accuracy, 91.77% F1, 89.08% AUC-ROC, outperforming classical baselines.

**Reference papers**:
- arXiv: 2604.22903 (Adaptive Hybrid Quantum-Classical Feature Fusion for Breast Cancer)

### Pattern 7: Tensor-Network Quantum Federated Learning

**Core idea**: Tensor-network frontends (MPS/TTN/MERA) compress local medical inputs
into compact latents, enabling small-qubit quantum post-aggregation refinement while
reducing MPC communication overhead simultaneously.

**Architecture**:
```
Clients: [MPS/TTN/MERA Frontend → Compressed Latent]
    ↓
Secure Aggregation (MPC)
    ↓
Quantum-Enhanced Processor (QEP): quantum-state embedding + observable readout
```

**Key findings**:
- **TTN+QEP** most balanced on PneumoniaMNIST
- QEP effect is frontend-dependent, not uniform
- Tensor-network compression: enables small-qubit quantum processing AND reduces MPC overhead

**Frontend selection**: MPS for 1D sequential, TTN for medical imaging, MERA for multi-scale.

**Reference papers**:
- arXiv: 2604.01616 (Tensor-Network Frontends for Privacy-Aware Federated Medical Diagnosis)

### Pattern 8: HQNN for Thermographic Medical Imaging

**Core idea**: Quantum variational layers within classical CNN for thermographic
breast cancer detection. Angle encoding maps thermal pixel intensities to rotations.

**Reference papers**:
- arXiv: 2604.16953 (HQNN for Breast Cancer Thermographic Classification, IEEE IBITeC 2025)

### Pattern 9: QSVM Feature Map Selection Protocol

**Core idea**: Systematic selection of quantum feature maps for QSVM classification in medical domains, with decision tree based on data dimensionality, non-linearity needs, and NISQ constraints.

**Feature Map Taxonomy**:

| Encoding | Qubits | Depth | Best For |
|----------|--------|-------|----------|
| Angle Encoding | n = features | O(1) | Normalized features, <20 dims |
| Amplitude Encoding | n = log₂(features) | O(N) | High-dimensional (images, genes) |
| ZZFeatureMap | n = features | O(reps × n²) | Non-linear separable, entanglement needed |
| IQPFeatureMap | n = features | O(n²) | Theoretical advantage studies |

**Selection Decision Tree**:
```
Data dimensionality?
├── Low (<20) → Angle Encoding or ZZFeatureMap
├── Medium (20-100) → PCA → Angle Encoding, or ZZFeatureMap on top components
└── High (>100) → Amplitude Encoding or CNN features → Angle Encoding

Non-linearity needed?
├── Yes → ZZFeatureMap (reps=2) or IQPFeatureMap
└── No → Angle Encoding

NISQ constraints?
├── Limited coherence → Angle Encoding (shallowest)
├── Moderate → ZZFeatureMap (reps=1)
└── Simulator → Full expressivity
```

**Evaluation Protocol**:
1. **Expressivity**: Kernel Target Alignment KTA = Tr(K_target · K_quantum) / (||K_target|| · ||K_quantum||)
2. **Generalization**: Quantum kernel condition number κ(K) — lower is better
3. **Accuracy**: Cross-validate on train set
4. **Advantage check**: Compare quantum kernel SVM vs classical RBF kernel SVM

**Medical Domain Specifics**:
- Thermographic: CNN features (512-d) → PCA (16-d) → ZZFeatureMap(reps=2)
- X-ray: PCA (8-d) → Angle Encoding
- Clinical tabular: Standardize → Amplitude Encoding
- Key insight: Feature map choice has larger impact than circuit ansatz — always benchmark multiple encodings

**Reference papers**:
- arXiv: 2506.03272 (Investigating Quantum Feature Maps in QSVM for Lung Cancer Classification)
- arXiv: 2505.20804 (QML in Healthcare: Evaluating QNN and QSVM Models)

### Pattern 10: Multi-VQC Ensemble for Imbalanced Healthcare Data

**Core idea**: Train multiple VQCs with different initializations and aggregate predictions to handle class imbalance in medical classification, where traditional models fail on minority classes.

**Workflow**:
```
Medical data (imbalanced) → Classical preprocessing
    → [VQC_1, VQC_2, ..., VQC_n] (different initializations/encodings)
    → Weighted ensemble aggregation → Diagnosis
```

**Key design choices**:
- Each VQC uses different random initialization or slightly different encoding
- Ensemble weights can be optimized on validation set (focus on minority class recall)
- More robust than single VQC which may converge to suboptimal local minima
- Particularly effective when minority class prevalence < 20%

**Reference papers**:
- arXiv: 2505.20797 (Multi-VQC: A Novel QML Approach for Enhancing Healthcare Classification)
- arXiv: 2505.14716 (Hybrid Quantum Classical Pipeline for X-Ray Based Fracture Diagnosis)

## Implementation Checklist

When researching or implementing quantum medical applications:

1. **Data characteristics**: Imbalanced datasets common in medical domain — use class weighting, SMOTE, or focal loss
2. **Encoding choice**: Match encoding strategy to data type (continuous → angle/CV, binary → basis)
3. **Circuit depth**: Keep within NISQ limits (≤20 layers) to avoid noise degradation
4. **Baseline comparison**: Always compare against classical counterparts (SVM, RF, CNN)
5. **Metric selection**: Use medically relevant metrics — sensitivity, specificity, AUC-ROC, not just accuracy
6. **Dataset size**: Quantum advantage more likely with small datasets (<10K samples) where classical models overfit
7. **Hardware awareness**: Specify target backend (simulator, IBM, IonQ, photonic) and noise model

## Common Pitfalls

- **Data encoding bottleneck**: Loading classical medical data into quantum states can erase any quantum advantage (the "input problem")
- **Barren plateaus**: Deep variational circuits suffer from vanishing gradients — use shallow circuits or layerwise training
- **Overclaiming**: Many papers show advantage only on toy datasets; real medical data is much noisier
- **Reproducibility**: Quantum simulators vs. real hardware show vastly different results
- **Class imbalance**: Medical datasets are typically heavily imbalanced — must address explicitly

## Search Queries for New Papers

### Sub-class taxonomy reference
For detailed sub-class classification and decision rules, see [references/quantum-medical-taxonomy.md](references/quantum-medical-taxonomy.md).

**Before creating new medicine+quantum skills**: Check against the 4 established sub-classes above. New papers should extend existing patterns or add genuinely new sub-classes only.

### Pattern 11: Entanglement-Based PET Imaging (J-PET)
**Core idea**: Exploit quantum entanglement of polarization between annihilation photons as an additional diagnostic channel beyond standard PET uptake.
- Plastic scintillators detect via Compton scattering → simultaneous position + time + polarization
- Entanglement degree = correlation of polarization planes between coincidence pairs
- First in vivo human imaging (2606.29421): liver/spleen entanglement between maximally entangled and separable
- PET with entangled Compton events (2606.25804): 20% higher SBR, up to 10% sensitivity increase

### Pattern 12: Parameter-Efficient CV Photonic QNNs for Edge Medical AI
**Core idea**: Room-temperature photonic quantum computing for edge medical classification with minimal parameters.
- Simplified Phi-D-U1 layer cuts parameters 40-45% vs full Killoran layer
- 18 parameters achieve 100% calibrated accuracy for oral cancer detection (2606.28252)
- Mitigates barren plateaus: gradient variance improved ~58 orders of magnitude
- MobileNetV1 + PCA 16-dim + CV-QNN pipeline

### Pattern 13: Compression-Driven Anomaly Detection via Quantum Autoencoders
**Core idea**: QAEs trained on normal data detect anomalies via incompressibility.
- Angle encoding + variational encoder-decoder with trash qubits (2606.27411)
- ROC-AUC ~0.95 slice-level, ~0.813 patch-level on brain MRI
- Encoder-decoder asymmetry: detection from structured compression, not decoder expressivity
- Spatially localized anomaly heatmaps aligned with tumorous regions

### Pattern 14: Quantum Ophthalmology
**Core idea**: Four complementary quantum approaches to ophthalmic imaging (2606.19238):
1. Photon-limited retinal imaging (OCT + single-photon detection)
2. Correlation-based imaging (ghost imaging for low-light/scattering)
3. Nanoscale optical probes (quantum dots for contrast/targeting)
4. Quantum-limited visual perception (single-photon level structured light)

```
site:arxiv.org quantum machine learning medical imaging
site:arxiv.org quantum neural network diagnosis
site:arxiv.org quantum drug discovery
site:arxiv.org quantum clinical trial
site:arxiv.org quantum kernel medical
site:arxiv.org continuous variable quantum neural network biomedical
```