Why Data Structure Matters for Quantum Machine Learning in Healthcare?

Abstract

The potential advantages of quantum machine learning (QML) over classical methods in clinical decision-support systems remain an active research area; yet because most QML evaluations are conducted on a single dataset, how the success of the quantum approach depends on data structure has not been adequately examined. Here we evaluate two structurally opposite health datasets - Wisconsin Breast Cancer Diagnostic for binary classification and Palechor Obesity Estimation for seven-class classification - under an identical cross-validation protocol, testing 35 model variants in total (16 on WBCD and 19 on the obesity dataset). For the multi-class task we introduce an original hybrid architecture (Q-Hybrid-Q3-Plus) that combines Angle and Amplitude encodings in parallel channels, and quantify performance through effect size (Cohen's d). Classical models were found to be statistically superior on both datasets (one-tailed Wilcoxon p=0.0312), but the magnitude of this superiority depends strongly on data structure: on WBCD the 5.05-percentage-point accuracy gap (d=1.70) widens to 14.68 points (d=5.13) on the obesity dataset. Our proposed Q-Hybrid-Q3-Plus achieves roughly 17 points of improvement over pure quantum baselines. The high parameter efficiency of quantum models - reaching 92.97% accuracy with only 54 parameters on WBCD - hints at practical benefits that may grow as NISQ hardware matures, while the cross-context evaluation protocol offers a reusable framework for further research. © 2026 IEEE.