HybridQMedKD

A lightweight hybrid quantum-classical framework for biomedical tabular binary classification under low-resource simulation.

Objective

We study whether a small quantum layer is useful for low-dimensional biomedical tabular classification when combined with:

• Classical teacher-student knowledge distillation
• PCA-based dimensionality reduction
• Different quantum-layer insertion positions (front / middle / tail)

Core Research Questions

1. Does knowledge distillation improve hybrid quantum-classical student models?
2. Where should the quantum layer be inserted for best performance-cost trade-off?
3. Under low-resource simulation, is the hybrid student worth the extra cost?

Dataset

• Primary: Breast Cancer Wisconsin (Diagnostic) — 569 samples, 30 continuous features, binary classification, no missing values
• Optional: any low-dimensional biomedical tabular binary classification dataset

Models

ID	Model	Distillation	Quantum Layer
G1	Teacher-Classic	No	No
G2	Student-Classic	No	No
G3	Student-Classic-KD	Yes	No
G4	Student-Hybrid-NoKD	No	Yes
G5	Student-Hybrid-KD	Yes	Yes

Metrics

• AUC, F1, Accuracy, MCC
• Training time, Inference time
• Qubit count, Circuit depth

Ablation

• Quantum layer position: front / middle / tail
• PCA dimensions: 4 / 6 / 8
• Distillation temperature T: 2 / 4
• Distillation alpha: 0.3 / 0.5 / 0.7

Project Structure

HybridQMedKD/
├── configs/
├── data/raw/
├── src/
│   ├── datasets/
│   ├── models/
│   ├── losses/
│   ├── trainers/
│   └── utils/
├── outputs/
└── scripts/

Quick Start

pip install -r requirements.txt
python src/main.py

Notes

• Keep quantum circuits shallow (4 qubits, 1-2 layers) for local simulation feasibility
• PCA / feature selection is applied within each training fold to prevent data leakage
• Goal is practical design insight, not quantum advantage claims