A Hybrid Quantum-Classical Machine Learning Framework for Early Parkinson’s Disease Detection from Vocal Biomarkers

Abstract

Phonation changes early in the progression of Parkinson's disease (PD), supporting sustained-vowel voice recordings as an appealing low-cost screening medium. It remains an open question, and one often over-claimed, whether near-term quantum machine learning is greater than or equal to classical models in this small-sample clinical regime. We introduce a controlled-leakage benchmark of classical and quantum models on the standard 195-recording Oxford PD dysphonia dataset (22 acoustic features; 147 PD and 48 healthy phonations from 32 speakers). For a relevant clinical metric (macro-F1 0.701, ROC-AUC 0.861, Matthews correlation coefficient (MCC) 0.449), a four-qubit variational quantum classifier (VQC) matched or outperformed the strongest classical baseline (a multilayer perceptron: macro-F1 0.638, ROC-AUC 0.825), with much weaker performance from a quantum-kernel support vector machine [41]. The naive record-wise split inflated the macro-F1 by 0.156 on average (tree ensembles were most inflated, up to 0.266; and the VQC and the perceptron were almost leakage-immune). Quantum models maintained discriminative power (AUC-ROC ≈ 0.86) under simulated depolarizing and bit-flip noise up to p=0.1, all circuits assessed resided in a barren plateau-free region suitable for training, and classical versus quantum feature-attribution rankings were negatively correlated (Spearman ρ =−0.431), posing caution for clinical interpretability. Performance differences between the best models were significant at a per-sample (McNemar p = 0.0012) but not across- fold (Wilcoxon p = 0.44) level, so quantum advantage is reported suggestively rather than definitively.