An analytics framework for privacy-preserving federated healthcare learning with quantum-inspired perturbation

Federated learning (FL) and differential privacy (DP) collectively provide a robust framework for preserving privacy in healthcare data analytics. This study extends this architecture by introducing quantum-inspired gradient perturbation and examining the impact of quantum-inspired noise models and their integration order within privacy-preserving federated learning systems. A comprehensive evaluation of federated averaging (FedAvg) and federated proximal (FedProx) optimization strategies is conducted under both independent and identically distributed (IID) and non-independent and identically distributed (non-IID) data settings. A control-variate aggregator (SCAFFOLD) is also evaluated on the core non-IID setting, and the core experiments are run across three random seeds with pairwise paired t-tests for statistical significance. The findings show that all approaches achieve high classification accuracy and macro-F1 performance while preserving privacy guarantees; across seeds, noise ordering and aggregator choice produce no statistically significant difference in final accuracy (all p ≥ 0.33), confirming that the privacy-enhancing tiers preserve utility at a fixed budget. The proposed framework is further compared with classical Gaussian and Laplacian perturbation methods, and controlled ablation experiments confirm the effectiveness of the different configurations. Results indicate that quantum-inspired noise produces smoother convergence behaviour during early communication rounds while maintaining strong predictive performance. An extended evaluation on a lightweight convolutional network with two further imaging modalities (PathMNIST, OrganAMNIST), a four-qubit entangled circuit, and active privacy attacks shows that pre-DP perturbation increases gradient-inversion error by approximately 7.4 × over DP-SGD alone; the budget-neutral quantum-inspired tier matches this gain at equivalent accuracy, its distinctive contribution being a statistically distinct noise distribution rather than a larger reconstruction error, while membership-inference advantage remains near chance. The evaluation includes privacy accounting, convergence stability, computational overhead, accuracy, loss, macro-F1 score, and confusion matrix analysis.