Adaptive quantum kernel selection via leakage-free stacking for clinical diagnostics on NISQ hardware
摘要
Quantum kernel methods map clinical features into exponentially large Hilbert spaces where overlapping biological markers can become more separable than in fixed-dimensional classical feature spaces, but existing work evaluates single quantum feature maps, ignores barren-plateau failure modes, and relies on single train–test splits vulnerable to data leakage. Classical diagnostics for Parkinson’s disease, breast cancer, and diabetes remain limited by the Specificity–Recall trade-off that fixed-dimensional kernels impose on overlapping biomarker distributions. We propose an adaptive hybrid quantum framework routing clinical data through three distinct quantum feature maps, namely Angle, Amplitude, and ZZ-entanglement, computing fidelity-based Gram matrices for Quantum SVM and Quantum KNN classifiers. A Logistic Regression meta-learner, trained on strictly out-of-fold predictions from nested cross-validation (5-fold inner, 10-fold outer), learns which quantum kernel generalizes on each dataset and suppresses those that do not. Evaluated on Parkinson’s (195 patients), Breast Cancer (569), and Diabetes (768) with 1,000-iteration bootstrapping, the ensemble raised Parkinson’s Specificity from 0.585 to 0.813 (