Quantum-Inspired KANs Boost Privacy in Federated Biosignal Learning

Electrocardiogram (ECG) recordings are highly sensitive biomedical data, which creates significant privacy barriers for hospitals and wearable device manufacturers when attempting to share raw signals for centralized AI model training. Federated learning offers a solution by enabling collaborative model training while keeping the raw biosignal data securely at its source, thus addressing privacy concerns. However, federated ECG classification still faces challenges such as limited client-side samples, imbalanced arrhythmia labels, and non-independent and identically distributed (non-IID) data across different clients. These constraints necessitate classifiers that are both communication-efficient and resilient to variations in data distribution across clients. In this research, a Hybrid Quantum-inspired Kolmogorov-Arnold Network (HQKAN) was evaluated against a traditional multilayer perceptron (MLP) for arrhythmia classification. The evaluation was conducted on the MIT-BIH and INCART datasets under a federated averaging (FedAvg) setup, across various client configurations. The results demonstrated that HQKAN consistently improved most aggregate and minority-class metrics. Crucially, HQKAN achieved these performance gains while using substantially fewer trainable parameters and significantly reducing communication costs compared to the MLP baseline. For instance, on the MIT-BIH dataset, HQKAN reduced parameters by 37.35% and communication cost by 24.89%. Similar reductions were observed on the INCART dataset. These findings position HQKAN as a compact, communication-efficient, and robust alternative for privacy-aware federated learning applications involving sensitive biosignal data.