Geometric and Quantum Kernel Methods for Predicting Skeletal Muscle Outcomes in chronic obstructive pulmonary disease

TL;DR

This study compares classical and quantum kernel methods for predicting muscle outcomes in COPD, demonstrating quantum approaches can outperform classical models in small datasets.

Contribution

It introduces structured nonlinear low-data strategies, including quantum-kernel regression, for transparent and robust small-sample biomedical predictions.

Findings

Quantum-kernel ridge regression achieved the best performance (RMSE 4.41 mg, R2 0.62).

SPD features improved prediction over classical ridge regression.

Screening evaluation reached ROC-AUC 0.91 for low muscle weight.

Abstract

Quantum methods are increasingly proposed for healthcare, but translational biomarker studies demand transparent benchmarking and robust small-dataset evaluation. We analysed a preclinical COPD cohort of 213 animals with blood and bronchoalveolar-lavage biomarkers to predict tibialis anterior muscle weight, specific force, and muscle quality. We benchmarked tuned classical models against two structured nonlinear low-data strategies: geometry-aware symmetric positive definite (SPD) descriptors, in which training-only clustering maps each subject to Stein-divergence distances from representative prototypes and optional unlabeled synthetic SPD interpolation stabilises prototype discovery; and quantum-kernel regression, including a clustered Nystrom-style feature map that compresses each subject into similarities to a small set of training-derived centres. By replacing full pairwise structure with compact prototype- and centre-based summaries, these steps regularise learning and preserve interpretability in a small-sample setting. Across five outer folds, quantum-kernel ridge regression using four interpretable inputs achieved the best muscle-weight performance (RMSE 4.41 mg; R2 0.62), outperforming a matched compact classical baseline (4.68 mg; R2 0.56). Biomarker-only SPD features also improved over ridge regression (4.55 versus 4.79 mg), and screening evaluation reached ROC-AUC 0.91 for low muscle weight.