A Pilot Ablation Study of Quantum-Layer Placement in Hybrid Networks for Biomedical Tabular Classification

Read full text Summarise with VIRAT
Authors: Xiaoming Liu
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Hybrid quantum-classical architectures are easy to overclaim on small biomedical benchmarks, where strong classical baselines and low statistical power can make minor differences look decisive. We revisit a lightweight teacher-student pipeline with PCA compression, a compact classical student, a 4-qubit parameterized quantum layer, and knowledge distillation (KD) for biomedical tabular classification. We rerun the main Breast Cancer Wisconsin Diagnostic (WDBC) comparison with 10-fold cross-validation, retain the original WDBC ablations, and add companion Pima Diabetes and Heart Statlog benchmarks plus pilot cross-dataset placement checks. On WDBC, the tail-placement Hybrid-KD model matches but does not reliably outperform the classical student (AUC 0.9926 ±0.0095 vs. 0.9947 ±0.0081), and the hybrid cost remains in the hundreds of seconds per fold rather than sub-second. The placement result also needs a narrower interpretation than our original draft: front insertion is fragile in this implementation because direct PCA-to-qubit mapping induces heavy Tanh saturation and removes the learnable pre-quantum projection, whereas middle and tail remain viable. Pima and Heart show only modest, statistically uncertain hybrid gains under much higher simulation cost. We therefore present the study as a pilot ablation of placement and input conditioning, not as definitive evidence that lightweight hybrid models outperform compact classical baselines.

Fields of Study & Subjects

Quantum Computing Algorithms and Architecture Machine Learning and Data Classification Artificial Intelligence in Healthcare

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
30 March 2026
Language
EN
Document type
Preprint