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; En-Jui Kuo; Hsi-Sheng Goan

Source-linked AI summary

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, En-Jui Kuo, Hsi-Sheng Goan

arXiv:2608.13914v1Aug 14, 2026cs.LGcs.AIcs.DCcs.ETquant-ph

TL;DR

Federated ECG classification must handle privacy-sensitive data alongside limited samples, class imbalance, heterogeneous clients, and communication constraints. This paper evaluates HQKAN against an MLP under FedAvg, finding stronger aggregate and minority-class performance with substantially lower parameter and communication costs across MIT-BIH and INCART.

Problem
Federated ECG classification requires privacy-preserving learning that remains effective under limited samples, class imbalance, client heterogeneity, and communication constraints.
Method
The study benchmarks HQKAN against an MLP under matched FedAvg budgets across IID and label-skewed partitions of MIT-BIH and INCART.
Results
37.35% fewer trainable parameters and 24.89% lower communication cost on MIT-BIH, with corresponding reductions of 44.81% and 36.41% on INCART, accompany stronger reported performance.
Takeaways & Limitations
HQKAN is a compact, communication-efficient alternative to MLP for federated ECG classification across the evaluated datasets and client heterogeneity settings.

Abstract

from arXiv · show

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.

I. INTRODUCTION

The introduction frames federated ECG classification as a privacy-preserving setting constrained by communication, client heterogeneity, and class imbalance. It presents HQKAN as a compact alternative to an MLP and evaluates both under matched federated conditions across MIT-BIH and INCART.

Motivation: ECG recordings are privacy-sensitive and distributed across patients, devices, and institutions, motivating federated learning that keeps data local while exchanging model updates.Federated ECG learning makes communication, client heterogeneity, and class imbalance central design constraints.
Method: QKANs use DARUAN edge functions as learnable activations and require far fewer trainable parameters than conventional MLPs.HQKAN extends QKAN with a fully connected encoder and decoder surrounding a QKAN latent feature processor.
Method: HQKAN uses a fully connected encoder and decoder with a QKAN latent feature processor, forming an autoencoder-like architecture.This hybrid architecture was first introduced in [13].
Evaluation: Both models are evaluated with identical client partitions and training budgets across K ∈{8, 16, 32} clients on MIT-BIH and INCART.MIT-BIH uses five ANSI/AAMI heartbeat classes, while INCART uses three ANSI/AAMI classes.
Contribution: 11,581 and 15,147 trainable parameters are used by HQKAN on MIT-BIH and INCART, compared with 18,485 and 27,443 for the MLP baseline, respectively.The parameter counts are reported for the five-class MIT-BIH and three-class INCART evaluations.

II. RELATED WORK

Related work frames federated learning as a privacy-aligned approach for distributed biomedical signal modeling, while highlighting non-IID data, class imbalance, and communication constraints in federated ECG. It also situates QKANs as quantum-inspired alternatives using trainable single-qubit edge functions instead of fixed spline activations.

Federated biomedical learning: Federated learning keeps biomedical data on-device and shares model updates, addressing privacy requirements for signals including ECG, electroencephalogram, and electromyography.Applying federated learning to ECG also involves non-IID data, class imbalance, and constrained communication.
Conventional model backbones: Conventional federated ECG efforts use backbones such as convolutional neural networks, long short-term memory networks, and residual networks.The cited conventional architectures include CNNs, LSTMs, and ResNets.
Quantum-inspired networks: QKANs use data re-uploading and trainable single-qubit circuits as edge functions to replace fixed spline activations in Kolmogorov–Arnold networks.Reported QKAN applications include forecasting, traffic-matrix modeling, quantum circuit generation, and language modeling.

III. METHODS · A. Dataset and Preprocessing

The study uses preprocessed MIT-BIH ECG beats, represented as normalized, zero-padded single-lead waveforms and divided into five arrhythmia classes. A seeded, label-stratified split creates training and validation subsets.

A. Dataset and Preprocessing: MIT-BIH Arrhythmia is the primary benchmark, with each heartbeat encoded as a 187-sample single-lead waveform.Waveforms are normalized to [0, 1] and zero-padded when necessary.
A. Dataset and Preprocessing: Five classes categorize beats as non-ectopic (N), supraventricular ectopic (S), ventricular ectopic (V), fusion (F), or unknown (Q).
A. Dataset and Preprocessing: 70,043 training and 17,511 validation beats result from a seeded label-stratified 80/20 split of the original training set.

B. Federated Setup

The federated setup uses FLamby with FedAvg, full client participation, and a fixed 30-round training budget. Experiments vary classifier, client count, and IID or Dirichlet-based non-IID partitions while using imbalance-aware local optimization.

Federated optimization: FedAvg initializes each client from global parameters, performs E local epochs, and aggregates the returned models at every communication round.The pipeline is built on FLamby and uses FedAvg as its only aggregation method.
Experimental configurations: 30 rounds and 5 local epochs are used for every configuration with full client participation in each communication round.Configurations vary HQKAN or MLP, K ∈ {8, 16, 32}, and IID or non-IID data partitions.
Data partitioning: IID data are distributed round-robin after class-wise shuffling, whereas non-IID client label proportions follow a Dirichlet distribution.The main non-IID experiments use α = 0.3, with robustness evaluated over α ∈ {0.1, 0.3, 0.5, 1, 1000}.
Local training and selection: Inverse-frequency-weighted cross-entropy addresses AAMI class imbalance, with AdamW using a learning rate of 10−3, weight decay of 10−4, and batch size 128.Model selection retains the global checkpoint with the highest validation macroaveraged AUPRC over the 30 rounds.

C. Classifier

The two classifiers use identical dataset-specific inputs and differ only in architecture. HQKAN is a compact alternative to the MLP, using substantially fewer trainable parameters on MIT-BIH through its QKAN-based latent processor.

Both models use identical inputs within each dataset and differ only in classifier architecture.
187 is the flattened input dimension for MIT-BIH, where the MLP baseline uses ReLU activations and 18,485 trainable parameters.
11,581 trainable parameters gives HQKAN a 37.35% reduction versus the MLP baseline on MIT-BIH.HQKAN combines a fully connected encoder and decoder with a QKAN latent feature processor.
300 is the corresponding input dimension for INCART.

IV. RESULTS AND DISCUSSION

The evaluation uses aggregate and class-wise metrics to assess classification performance, with macro-F1 as the primary measure. Client label heterogeneity is quantified by mean Hellinger distance, where larger values indicate stronger non-IID skew.

Evaluation metrics: Macro-F1 is the key performance measure, complemented by AUROC and AUPRC.Performance is also reported with Cohen’s κ and the Brier score.
Evaluation metrics: Class-wise performance is evaluated using PPV, sensitivity, and specificity under a one-vs-rest convention.These metrics are reported alongside macro-F1, Cohen’s κ, and the Brier score.
Client heterogeneity: Mean Hellinger distance H(p, q) summarizes each client’s label-distribution difference from the global distribution q.The distance is defined over local distribution p and global distribution q, with H ∈ [0, 1].
Client heterogeneity: Larger Hellinger distance values indicate stronger non-IID skew across clients.Hellinger distance is used to summarize client label heterogeneity.

A. Aggregate Performance

HQKAN consistently outperforms the MLP baseline on aggregate ECG classification metrics across MIT-BIH and INCART under IID and non-IID federated partitions. Its advantage persists at higher client counts, including stronger macro-F1 and lower Brier scores at 32 clients.

A. Aggregate Performance: HQKAN consistently exceeds the MLP baseline on all reported aggregate metrics across both datasets under IID and non-IID partitions.Under non-IID settings, HQKAN achieves higher macro-F1 and κ and lower Brier scores across client counts.
A. Aggregate Performance: 0.013 to 0.054 is the increase in HQKAN’s IID macro-F1 advantage over MLP as MIT-BIH clients grow from 8 to 32.This pattern suggests greater robustness as federation becomes more distributed.
A. Aggregate Performance: 0.761 macro-F1 versus 0.698 for MLP at 32 clients on MIT-BIH demonstrates HQKAN’s aggregate advantage in the most distributed setting.HQKAN also reduces the Brier score from 0.121 to 0.094.
A. Aggregate Performance: 0.850 macro-F1 versus 0.838 for MLP at 32 clients on INCART confirms the same aggregate advantage across ECG benchmarks.HQKAN reduces the INCART Brier score from 0.040 to 0.032.

B. Minority-Class Performance

HQKAN achieved higher mean values on most minority-class measures across MIT-BIH and INCART, but class-wise trade-offs varied by metric and partition regime. Precision-recall curves showed the clearest model separation for class F on MIT-BIH and class S on INCART.

Cross-dataset overview: HQKAN achieved higher mean values on most minority-class PPV, sensitivity, and specificity measures across both datasets, with patterns varying by class, metric, and partition regime.Tables III and IV report these class-wise results at 32 clients.
MIT-BIH: 0.147, 0.077, and 0.073 were HQKAN’s non-IID mean sensitivity advantages over MLP for MIT-BIH classes S, V, and F, respectively.For class S, PPV was slightly lower (0.755 vs 0.768), whereas HQKAN had higher PPV for V and F.
INCART: 0.079 and 0.013 were HQKAN’s IID PPV and specificity advantages for INCART class S, while sensitivity was lower by 0.005.Under non-IID partitioning, PPV and specificity were higher by 0.083 and 0.001, while sensitivity was lower by 0.023.
INCART: 0.008 and 0.020 were HQKAN’s higher mean sensitivity values for INCART class V under IID and non-IID partitioning, respectively.For class S, higher PPV coincided with lower sensitivity under both partitions.
Threshold analysis: Model separation was most pronounced for class F on MIT-BIH and class S on INCART in the class-wise precision-recall curves.Figures 1 and 2 characterize the metric trade-offs across decision thresholds.

C. Robustness to Client Heterogeneity

At 32 clients across five Dirichlet concentrations, HQKAN consistently outperformed MLP in mean AUROC and AUPRC on both datasets. Its advantage widened under stronger label skew, especially for AUPRC, as MLP performance declined more sharply.

Cross-client heterogeneity: At 32 clients across five Dirichlet concentrations, HQKAN achieved higher mean AUROC and AUPRC than MLP at every tested concentration on both datasets.Hellinger distance summarizes client-label heterogeneity, with larger values indicating stronger label skew.
Cross-client heterogeneity: At the largest Hellinger distances, HQKAN’s margins over MLP widened, particularly for AUPRC.The largest Hellinger distances correspond to stronger label skew across clients.
Cross-client heterogeneity: MLP’s mean performance declined more sharply as client label distributions became more skewed, supporting HQKAN’s robustness.This pattern was observed across the tested heterogeneity conditions.

V. CONCLUSION

HQKAN is evaluated as a parameter-efficient federated ECG classifier against an MLP across datasets, client configurations, and data partitions. Across evaluated settings, it improves key aggregate metrics while maintaining stronger discrimination and precision-recall performance.

Evaluation setup: HQKAN was benchmarked against an MLP under matched FedAvg budgets with 8, 16, and 32 clients on MIT-BIH and INCART datasets.Evaluation covered IID and label-skewed partitions, including five-class MIT-BIH and three-class INCART classification.
Aggregate performance: HQKAN outperforms the MLP in mean macro-F1 and Cohen’s κ while achieving lower mean Brier scores across evaluated configurations.These results indicate improved aggregate classification quality and calibration under the tested federated settings.
Minority-class discrimination: HQKAN maintains higher AUROC and AUPRC across all evaluated levels than the MLP baseline.ROC and precision-recall curves for minority classes S, V, and F were averaged over five random seeds under IID and non-IID MIT-BIH partitions.
Per-class evaluation: At 32 clients and 5 local epochs, MIT-BIH per-class performance was reported as mean±STD over 5 seeds, with non-IID partitions using α = 0.3.The table specifies the evaluation condition for the per-class results.