Hybrid Quantum-inspired Kolmogorov-Arnold Networks for Privacy-Aware Federated Biosignal Learning

Chun-Hua Lin; Samuel Yen-Chi Chen; Yu-Chao Hsu; Kuo-Chung Peng; Jiun-Cheng Jiang; Chi-Sheng Chen; Tai-Yue Li; Nan-Yow Chen

cs.LGAug 14, 2026

Hybrid Quantum-inspired Kolmogorov-Arnold Networks for Privacy-Aware Federated Biosignal Learning

Authors: Chun-Hua Lin, Samuel Yen-Chi Chen, Yu-Chao Hsu, Kuo-Chung Peng, Jiun-Cheng Jiang, Chi-Sheng Chen, Tai-Yue Li, Nan-Yow Chen, +2 more

Organizations: Department of Physics and Center for Theoretical Physics, National Taiwan University, Taipei, Taiwan · National Center for High-Performance Computing, National Institutes of Applied Research, Hsinchu, Taiwan · Brookhaven National Laboratory, Upton, NY, USA · School of Electrical Engineering, Korea Advanced Institute of Science and Technology, Daejeon, Korea · Center for Quantum Science and Engineering, National Taiwan University, Taipei, Taiwan · Beth Israel Deaconess Medical Center, Harvard University, Boston, MA, USA · Department of Electrophysics, National Yang Ming Chiao Tung University, Hsinchu, Taiwan · Physics Division, National Center for Theoretical Sciences, Taipei, Taiwan

Abstract

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 due to limited client-side samples, imbalanced arrhythmia labels, and non-independent and identically distributed (non-IID) data across clients. These constraints require classifiers that are both communication-efficient and robust to cross-client distribution shifts. In this work, we evaluate a hybrid quantum-inspired Kolmogorov-Arnold network (HQKAN) against a multilayer perceptron (MLP) for five-class arrhythmia classification on the MIT-BIH dataset and three-class classification on the INCART dataset under federated averaging (FedAvg). Across multiple client configurations, HQKAN improves most aggregate and minority-class metrics while using 37.35% fewer trainable parameters and reducing communication cost by 24.89% on MIT-BIH; on INCART, it achieves corresponding reductions of 44.81% and 36.41%. These results indicate that HQKAN offers a compact, communication-efficient and robust alternative to the MLP baseline for privacy-aware federated learning on biosignal data.

Figures & tables

Fig. 1: ROC curves (a) and precision-recall curves (b) for the three minority classes (S, V, F) under IID and non-IID federated partitions on MIT-BIH, comparing the MLP baseline against HQKAN. Curves are averaged over five random seeds.

Fig. 2: ROC curves (a) and precision-recall curves (b) for the majority class N and minority classes S and V under IID and non-IID federated partitions on INCART, comparing the MLP baseline with HQKAN. Curves are averaged over five random seeds.

Setting	Clients	Model	macro-F1 $\uparrow$	$\kappa\uparrow$	Brier $\downarrow$
IID	8	MLP	0.726 $\pm$ 0.008	0.764 $\pm$ 0.011	0.127 $\pm$ 0.007
	8	HQKAN	0.739 $\pm$ 0.006	0.776 $\pm$ 0.008	0.119 $\pm$ 0.005
	16	MLP	0.672 $\pm$ 0.007	0.686 $\pm$ 0.013	0.178 $\pm$ 0.009
	16	HQKAN	0.716 $\pm$ 0.018	0.744 $\pm$ 0.025	0.137 $\pm$ 0.017
	32	MLP	0.633 $\pm$ 0.008	0.626 $\pm$ 0.013	0.232 $\pm$ 0.011
	32	HQKAN	0.687 $\pm$ 0.012	0.701 $\pm$ 0.017	0.164 $\pm$ 0.011

TABLE I: Performance of MIT-BIH dataset with 5 local epochs (mean

\pm

std over 5 seeds). Non-IID uses

\alpha=0.3

Setting	Clients	Model	macro-F1 $\uparrow$	$\kappa\uparrow$	Brier $\downarrow$
IID	8	MLP	0.815 $\pm$ 0.015	0.882 $\pm$ 0.013	0.047 $\pm$ 0.007
	8	HQKAN	0.838 $\pm$ 0.023	0.898 $\pm$ 0.017	0.040 $\pm$ 0.009
	16	MLP	0.764 $\pm$ 0.014	0.824 $\pm$ 0.017	0.071 $\pm$ 0.007
	16	HQKAN	0.806 $\pm$ 0.036	0.864 $\pm$ 0.036	0.054 $\pm$ 0.011
	32	MLP	0.743 $\pm$ 0.014	0.793 $\pm$ 0.026	0.090 $\pm$ 0.011
	32	HQKAN	0.773 $\pm$ 0.032	0.827 $\pm$ 0.039	0.068 $\pm$ 0.014

TABLE II: Performance of INCART dataset with 5 local epochs (mean

\pm

std over 5 seeds). Non-IID uses

\alpha=0.3

Setting	Class	Model	PPV	Sen	Spe
IID	S	MLP	0.282 $\pm$ 0.009	0.739 $\pm$ 0.018	0.951 $\pm$ 0.003
	S	HQKAN	0.334 $\pm$ 0.030	0.747 $\pm$ 0.014	0.961 $\pm$ 0.006
	V	MLP	0.669 $\pm$ 0.017	0.881 $\pm$ 0.003	0.969 $\pm$ 0.002
	V	HQKAN	0.705 $\pm$ 0.019	0.908 $\pm$ 0.012	0.973 $\pm$ 0.003
	F	MLP	0.147 $\pm$ 0.015	0.868 $\pm$ 0.007	0.962 $\pm$ 0.004
	F	HQKAN	0.239 $\pm$ 0.022	0.837 $\pm$ 0.014	0.980 $\pm$ 0.003

TABLE III: Per-class performance at 32 clients and 5 local epochs of MIT-BIH dataset (mean

\pm

std over 5 seeds). Non-IID uses

\alpha=0.3

Setting	Class	Model	PPV	Sen	Spe
IID	N	MLP	0.994 $\pm$ 0.000	0.945 $\pm$ 0.010	0.961 $\pm$ 0.003
	N	HQKAN	0.995 $\pm$ 0.001	0.956 $\pm$ 0.013	0.966 $\pm$ 0.004
	S	MLP	0.199 $\pm$ 0.025	0.838 $\pm$ 0.009	0.961 $\pm$ 0.008
	S	HQKAN	0.278 $\pm$ 0.076	0.833 $\pm$ 0.013	0.974 $\pm$ 0.009
	V	MLP	0.916 $\pm$ 0.014	0.966 $\pm$ 0.004	0.989 $\pm$ 0.002
	V	HQKAN	0.895 $\pm$ 0.038	0.974 $\pm$ 0.005	0.985 $\pm$ 0.006

TABLE IV: Per-class performance at 32 clients and 5 local epochs of INCART dataset (mean

\pm

std over 5 seeds). Non-IID uses

\alpha=0.3

Fig. 3: Robustness to client label heterogeneity at 32 clients and 5 local epochs of MIT-BIH dataset. The x-axis uses Hellinger distance to summarize label skew, with larger values indicating stronger non-IID partitions. The y-axis reports AUROC or AUPRC of different models. Each point corresponds to one Dirichlet concentration

\alpha

Fig. 4: Robustness to client label heterogeneity at 32 clients and 5 local epochs of INCART dataset. The x-axis uses Hellinger distance to summarize label skew, with larger values indicating stronger non-IID partitions. The y-axis reports AUROC or AUPRC of different models. Each point corresponds to one Dirichlet concentration

\alpha

Explore similar work

Towards Family-Grouped Hierarchical Federated Learning on Sub-5KB Models: A Feasibility Study of Privacy-Preserving ECG Monitoring for Ultra-Resource-Constrained Wearables
May 15, 2026Hangyu WuHierarchical Federated Learning Communication-Efficient Distributed Training
Shenzhen Coddie Technology co.,ltd, Shenzhen, 518000, China
BiFedKD: Bidirectional Federated Knowledge Distillation Framework for Non-IID and Long-Tailed ECG Monitoring
May 14, 2026Zixuan Shu, Tiancheng Cao, Hen-Wei HuangLong-Tail Learning Communication-Efficient Distributed Training
School of Electrical and Electronic Engineering, Nanyang Technological University, Republic of Singapore · Lee Kong Chian School of Medicine, Nanyang Technological University, Republic of Singapore
Privacy-Preserving Federated Autoencoder for ECG Anomaly Detection on Edge Devices
Jun 10, 2026Kaan Arda Akyol, Jakub Kacper Szeląg, Aydin Abadi +9Electrocardiogram Classification Reconstruction-Based Anomaly Detection
School of Computing, Newcastle University, Newcastle upon Tyne, United Kingdom

May 15, 2026J/K move · Enter open · S save

Towards Family-Grouped Hierarchical Federated Learning on Sub-5KB Models: A Feasibility Study of Privacy-Preserving ECG Monitoring for Ultra-Resource-Constrained Wearables

Hangyu Wu

Shenzhen Coddie Technology co.,ltd, Shenzhen, 518000, China

Hierarchical Federated Learning Communication-Efficient Distributed Training Electrocardiogram Classification Wearable Health Monitoring Federated Learning

Open paper AskPDF arXiv

Cardiovascular disease remains the leading cause of death worldwide, and early detection of arrhythmias through continuous ECG monitoring on wearable devices can prevent life-threatening events. Federated Learning (FL) enables privacy-preserving collaborative training by keeping raw ECG data on device, yet standard FL incurs prohibitive communication overhead and standard deep learning models cannot fit on ultra-low-power microcontrollers. We propose Family-Grouped Hierarchical Federated Learning (Family-FL), a three-tier architecture that uses the family as a natural privacy boundary for intra-family aggregation before global synchronization. We further design a hardware-constrained Tiny CNN-LSTM architecture with only 669 parameters, INT8-quantized to occupy merely 4.65KB Flash and 2.95KB RAM, meeting the constraints of STC32G12K128-class microcontrollers. Experiments on the MIT-BIH Arrhythmia Database (mean of 5 independent runs with different seeds) demonstrate that Family-FL reduces communication volume by 76.7% compared to FedAvg while maintaining comparable accuracy. Family-FL-Tiny achieves 91.9 +/- 1.2% accuracy with macro-F1 of 0.483 +/- 0.031, reducing total communication to 0.31% of FedAvg. The model achieves reliable ventricular arrhythmia detection (per-class F1 = 0.80), the most clinically critical abnormality for home-based preliminary screening. These results demonstrate the technical feasibility of privacy-preserving federated learning on ultra-resource-constrained microcontrollers through simulation-based evaluation. We honestly discuss limitations: no hardware deployment, single-dataset validation (MIT-BIH, 47 subjects), reduced rare-class sensitivity, and absence of formal differential privacy guarantees.

Abstract

Explore similar work

Towards Family-Grouped Hierarchical Federated Learning on Sub-5KB Models: A Feasibility Study of Privacy-Preserving ECG Monitoring for Ultra-Resource-Constrained Wearables

BiFedKD: Bidirectional Federated Knowledge Distillation Framework for Non-IID and Long-Tailed ECG Monitoring

Privacy-Preserving Federated Autoencoder for ECG Anomaly Detection on Edge Devices