Electrocardiogram (ECG) recordings are sensitive biomedical data, limiting the ability of hospitals and wearable devices to share raw signals for centralized model training. Federated learning addresses this practical privacy constraint by enabling collaborative model training while keeping raw biosignal data at their respective sources. However, federated ECG classification remains challenging du
Hybrid Quantum-inspired Kolmogorov-Arnold Networks (HQKAN) represent a parameter-efficient, communication-optimized alternative to traditional Multilayer Perceptrons (MLP) for privacy-aware federated biosignal learning, specifically addressing the challenges of training on sensitive Electrocardiogram (ECG) data.
Performance and Efficiency Gains Parameter Reduction: HQKAN uses 37.35% fewer trainable parameters on the MIT-BIH dataset and 44.81% fewer on the INCART dataset compared to MLP baselines. Communication Savings: The architecture reduces communication costs by 24.89% on MIT-BIH and 36.41% on INCART, critical for bandwidth-constrained federated environments. * Robustness: It achieves superior aggregate and minority-class metrics, including higher macro-F1 and Cohen’s κ scores, while maintaining lower Brier scores across heterogeneous, non-IID client distributions.
Architectural and Privacy Advantages Hybrid Design: HQKAN combines a fully connected encoder/decoder with a Quantum-inspired Kolmogorov-Arnold (QKAN) latent feature processor, enabling compact representation of complex nonlinear ECG patterns. Privacy Preservation: By keeping raw biosignal data local to hospitals and wearable devices, federated learning prevents the sharing of sensitive records, while HQKAN’s efficiency mitigates risks associated with gradient leakage through smaller model updates. * Clinical Applicability: This approach effectively handles class imbalance and limited client-side samples, making it suitable for real-world IoT healthcare deployments where device capabilities and data distributions vary significantly.
The paper addresses a practical barrier in biomedical machine learning: electrocardiogram (ECG) data are highly sensitive, so hospitals, clinics, and wearable devices often cannot pool raw signals into a centralized training repository. It positions federated learning as a way to train ECG classification models collaboratively while keeping raw biosignals local. The central challenge is that ECG classification in this distributed setting is difficult because of heterogeneous recording devices, variable signal quality, limited local data per site, and the need to communicate only constrained model updates rather than raw traces.
The proposed approach combines two ideas: Kolmogorov–Arnold Networks (KANs), which use learnable activation functions on network edges to offer a structured, potentially more parameter-efficient alternative to conventional MLPs, and quantum-inspired modeling or optimization techniques that are used to improve expressivity, training behavior, or feature representation without requiring full quantum hardware. The resulting hybrid architecture is embedded in a privacy-aware federated learning pipeline in which participating sites train on their own ECG data and share only model-level information, thereby reducing the risk of exposing raw patient records.
This work matters because it targets a realistic deployment scenario for clinical and wearable ECG analytics, where centralized data collection is often infeasible due to privacy, regulatory, and infrastructure constraints. More broadly, it explores how newer expressive architectures and quantum-inspired methods can be adapted to resource- and privacy-constrained biomedical federated learning, offering a potential path toward collaborative ECG model development that balances performance, communication efficiency, and data protection.