Atrial Fibrillation Detection with Arbitrary Leads via a Codebook-Based Reconstruction-Classification Framework

Hongtao Li; Jia Wei; Guoyao Li; Yuchen Lei; Guangnian Ma; Jia Xiao; Yuanjun Lai; Shuzhen Lv

cs.LGAug 19, 2026

Atrial Fibrillation Detection with Arbitrary Leads via a Codebook-Based Reconstruction-Classification Framework

Authors: Hongtao Li, Jia Wei, Guoyao Li, Yuchen Lei, Guangnian Ma, Jia Xiao, Yuanjun Lai, Shuzhen Lv,

Organizations: Information Center, The People’s Hospital of Baoan Shenzhen, Shenzhen, 518101, Guangdong, China · School of Computer Science and Engineering, South China University of Technology, Guangzhou, 510006, Guangdong, China · Department of Cardiology, The People’s Hospital of Baoan Shenzhen, Shenzhen, 518101, Guangdong, China · Institute for Advanced Study, Shenzhen University, Shenzhen, 518060, Guangdong, China

Abstract

Background and Objective: Reliable atrial fibrillation (AF) detection from electrocardiogram (ECG) signals remains challenging in real-world clinical settings due to variable lead configurations, cross-dataset domain shifts, and pervasive physiological and technical artifacts. So we develop a robust and generalizable deep learning model for accurate AF detection.\ Methods: We propose the Dual-Codebook Graph Collaborative Network (DCGCNet), a novel end-to-end vector-quantized variational autoencoder that jointly performs AF classification and ECG reconstruction. DCGCNet introduces two key components: (1) a Local-Global Contrastive Module for learning noise-invariant representations, and (2) an Adaptive Codebook Vector Quantizer that dynamically refines codebook prototypes to better align with input data distributions, thereby preventing codebook collapse and enhancing generalization.\ Results: DCGCNet achieves state-of-the-art performance in standard intra-dataset 12-lead evaluation and demonstrates exceptional cross-dataset generalization across seven diverse settings, consistently attaining AUC > 0.98 in all cases. Furthermore, it maintains high diagnostic accuracy under realistic noisy conditions, including baseline wander, powerline interference, and EMG artifacts.\ Conclusions: DCGCNet establishes a new benchmark for robust, generalizable, and noise-resilient AF detection, showing strong potential for deployment in real-world clinical environments.

Figures & tables

Figure 1: Evolution of AF detection frameworks and our motivation. (a) Fixed-Lead models are inflexible to varying inputs. (b) Arbitrary-Lead models suffer from information scarcity. (c) Sequential frameworks decouple tasks, causing error propagation. (d) Our parallel framework jointly optimizes both tasks for superior performance.

Figure 2: An end-to-end unified model integrating morphology-rhythm dual encoding, adaptive codebook quantization, local-global contrastive learning and heterogeneous graph interaction for multi-objective ECG analysis.

Figure 3: Experimental Protocol for Intra- and Cross-Dataset Evaluation of AF Detection Models.

Dataset	Normal	AF	Hz	Country	Leads
Chapman (Train)	6508	1416	500	China	12
CPSC2018 (Train)	1710	2215	500	China	12
Chapman (Test)	1617	364	500	China	12
CPSC2018 (Test)	423	531	500	China	12
AFDB	36718	22053	250	America	2
LTAF	187470	154590	128	America	2

Table 1: Summary of the segmented datasets used in this study. The number of “Normal” and “AF” samples are reported after 10-second segmentation.

Model	Chapman Lead I				Chapman Lead I+II+V1				CPSC2018 Lead I				CPSC2018 Lead I+II+V1
Model	PCC $\uparrow$	R 2 $\uparrow$	RMSE $\downarrow$	MAE $\downarrow$	PCC $\uparrow$	R 2 $\uparrow$	RMSE $\downarrow$	MAE $\downarrow$	PCC $\uparrow$	R 2 $\uparrow$	RMSE $\downarrow$	MAE $\downarrow$	PCC $\uparrow$	R 2 $\uparrow$	RMSE $\downarrow$	MAE $\downarrow$
ResCNN [ 34 ]	0.7828 (0.0012)	0.6424 (0.0016)	0.5983 (0.0014)	0.3173 (0.0012)	0.9349 (0.0006)	0.8881 (0.0012)	0.3402 (0.0020)	0.1643 (0.0016)	0.7519 (0.0013)	0.5933 (0.0053)	0.6406 (0.0042)	0.3387 (0.0016)	0.9132 (0.0011)	0.8537 (0.0021)	0.3853 (0.0029)	0.1800 (0.0022)
GRU [ 21 ]	0.7770 (0.0010)	0.6328 (0.0018)	0.6064 (0.0015)	0.3257 (0.0023)	0.9336 (0.0011)	0.8844 (0.0021)	0.3460 (0.0033)	0.1768 (0.0021)	0.7541 (0.0047)	0.6025 (0.0048)	0.6332 (0.0038)	0.3397 (0.0025)	0.9115 (0.0006)	0.8521 (0.0014)	0.3875 (0.0019)	0.1910 (0.0082)
EKGAN [ 20 ]	0.7272 (0.0011)	0.5342 (0.0122)	0.6801 (0.0089)	0.3714 (0.0043)	0.9139 (0.0016)	0.8320 (0.0042)	0.4084 (0.0050)	0.1983 (0.0047)	0.7309 (0.0025)	0.5177 (0.0202)	0.6941 (0.0144)	0.3652 (0.0048)	0.9190 (0.0026)	0.8441 (0.0046)	0.3947 (0.0058)	0.1826 (0.0043)
ESN [ 18 ]	0.6977 (0.0003)	0.5276 (0.0014)	0.6849 (0.0010)	0.3624 (0.0002)	0.9099 (0.0002)	0.8261 (0.0003)	0.4156 (0.0004)	0.1921 (0.0003)	0.6617 (0.0012)	0.4825 (0.0010)	0.7191 (0.0007)	0.3754 (0.0006)	0.8893 (0.0004)	0.7997 (0.0014)	0.4474 (0.0015)	0.1988 (0.0010)
DCGCNet(Ours)	0.7865 (0.0015)	0.6407 (0.0025)	0.5974 (0.0021)	0.3086 (0.0009)	0.9376 (0.0009)	0.8811 (0.0019)	0.3437 (0.0027)	0.1530 (0.0018)	0.7583 (0.0015)	0.6012 (0.0058)	0.6313 (0.0046)	0.3259 (0.0014)	0.9158 (0.0010)	0.8448 (0.0021)	0.3938 (0.0027)	0.1714 (0.0011)

Table 2: Performance Comparison of Different Models for Lead Reconstruction Performance on Chapman and CPSC2018 Datasets (mean

\pm

std).

\uparrow

: higher is better;

\downarrow

: lower is better. Bold: best result; underlined: second-best result.

Model	Chapman Dataset					CPSC2018 Dataset
Model	AUC	ACC	F1	Pre	Recall	AUC	ACC	F1	Pre	Recall
DNN [ 46 ]	0.9790 (0.0027)	0.9332 (0.0033)	0.8840 (0.0053)	0.9016 (0.0097)	0.8692 (0.0080)	0.9841 (0.0019)	0.9306 (0.0099)	0.9301 (0.0097)	0.9293 (0.0090)	0.9330 (0.0082)
DenseRNN [ 24 ]	0.9974 (0.0003)	0.9694 (0.0047)	0.9459 (0.0089)	0.9759 (0.0034)	0.9217 (0.0137)	0.9965 (0.0005)	0.9157 (0.0201)	0.9156 (0.0201)	0.9198 (0.0158)	0.9238 (0.0181)
ConvRGNN [ 42 ]	0.9627 (0.0048)	0.9318 (0.0040)	0.8832 (0.0095)	0.8948 (0.0086)	0.8739 (0.0201)	0.9545 (0.0091)	0.8889 (0.0109)	0.8875 (0.0112)	0.8878 (0.0113)	0.8881 (0.0120)
MSGformer [ 19 ]	0.9863 (0.0033)	0.9556 (0.0052)	0.9246 (0.0102)	0.9322 (0.0077)	0.9183 (0.0194)	0.9962 (0.0004)	0.9730 (0.0044)	0.9727 (0.0043)	0.9721 (0.0049)	0.9735 (0.0036)
ECGMamba [ 43 ]	0.9910 (0.0012)	0.9637 (0.0018)	0.9385 (0.0034)	0.9458 (0.0055)	0.9317 (0.0076)	0.9931 (0.0014)	0.9602 (0.0036)	0.9597 (0.0036)	0.9595 (0.0041)	0.9607 (0.0039)

Table 3: Performance Comparison of Different Models for Twelve-Lead Atrial Fibrillation Detection on Chapman and CPSC2018 Datasets (mean

\pm

std). Bold: best result

Figure 4: Cross-dataset generalization performance radar plots on AF detection comparing DCGCNet against baseline methods under three input strategies (fixed, arbitrary, reconstruction). (a) Models trained on CPSC2018 and evaluated on six external test sets. (b) Models trained on Chapman and tested on six datasets. Metrics shown include AUC, Accuracy, F1-score, Precision, and Recall.

Figure 5: Robustness of DCGCNet and RawECGNet on AF detection under increasing levels of ECG artifacts. AUC performance is plotted across four training–testing configurations as a function of noise intensity, defined as the ratio of artifact amplitude to original ECG amplitude (Ratio). Each curve shows the mean AUC over five independent runs, with error bars indicating the corresponding standard deviation. The clinically relevant 10 dB SNR level ( $\textit{Ratio}\approx 0.32$ )—representative of typical ECG interference in wearable settings—is highlighted by white markers.

Setting	Rec. (PCC)		Cls. (AUC)
Setting	Chapman	CPSC2018	Chapman	CPSC2018
DCGCNet	0.7865	0.7583	0.9998	0.9991
w/o rec	-	-	0.9883	0.9749
w/o cls	0.7751	0.7553	-	-
w/o rhythm	0.7812	0.7521	0.9823	0.9754
w/o morph	0.7205	0.6957	0.9986	0.9972

Table 4: Ablation study on intra-dataset: Lead I ECG reconstruction (PCC) and AF detection (AUC).

Setting	CPSC2018(12 Leads)	AFDB(2 Leads)	SPHDB(1 Lead)
DCGCNet	0.9967	0.9951	0.9947
w/o AC	0.9792	0.9715	0.9624
w/o ACVQ	0.9658	0.9543	0.9431

Table 5: Ablation study on cross-dataset generalization for AF detection (trained on Chapman).

	Chapman			CPSC2018
Setting	BLW	PLI	EMG	BLW	PLI	EMG
DCGCNet	0.9941	0.9952	0.9889	0.9863	0.9914	0.9807
w/o ACVQ	0.9624	0.9681	0.9432	0.9415	0.9523	0.9218
w/o LGCM_L	0.9832	0.9847	0.9603	0.9684	0.9721	0.9456
w/o LGCM_G	0.9756	0.9802	0.9714	0.9592	0.9667	0.9523
w/o LGCM	0.9518	0.9584	0.9276	0.9237	0.9341	0.9024

Table 6: Ablation study on robustness to common ECG noises (ratio=0.333) —BLW, PLI, and EMG—within the intra-dataset.

Figure 6: Codebook usage patterns across training and testing domains reveal consistent, label-specific prototypes, demonstrating the generalizability of learned discrete representations. (a) Model trained on CPSC2018 and tested on CPSC2018; (b) Same model tested on Chapman (out-of-domain). (c) Model trained on Chapman and tested on Chapman; (d) Same model tested on CPSC2018 (out-of-domain).

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A.1: 12-lead ECG reconstruction results generated by our proposed DCGCNet. (a) Reconstruction using the combination of Leads I, II, and V1 as input. (b) Reconstruction using only Lead I as input. In both subplots, the ground-truth 12-lead ECG signals are plotted as solid black lines, while the DCGCNet-predicted reconstructions are shown as dashed red lines.

Figure A.2: Effect of clinically relevant additive noise on the original 12-lead ECG signal. Three common artifacts—baseline wander (caused by respiration or motion, simulated as low-frequency sinusoids ¡0.3 Hz), powerline interference (from mains electricity, modeled as a 50 Hz sine wave with random phase), and electromyographic (EMG) artifact (due to muscle activity, generated as bandpass-filtered white noise in 20–250 Hz)—are injected at intensity ratios of 0.1–0.5 relative to the ECG’s peak-to-peak amplitude, ensuring physiologically realistic corruption.

Train	Test	Input	AUC	ACC	F1	Precision	Recall	Method	Other
AFDB	AFDB	30s	99.50	97.10	–	–	97.00	Wavelet + SVM [ 2 ]	2-fold CV
AFDB	AFDB	30 beats	–	97.60	97.10	–	98.00	Peak + RF [ 17 ]	4-fold CV
AFDB	AFDB	1.2s	–	99.23	–	–	99.41	Wavelet + CNN [ 15 ]	4:1 split
AFDB	AFDB	30s	–	97.80	–	–	98.98	CNN + LSTM [ 1 ]	5-fold CV
AFDB	AFDB	30s	–	93.00	92.64	–	89.83	Siamese + Few-shot TL [ 39 ]	5-fold CV
AFDB	AFDB	15s	–	99.11	97.00	99.46	95.40	Spatio-temporal Siamese [ 28 ]	8:1:1 split

Appendix

Table A.1: Comparison of AF detection performance with state-of-the-art methods across multiple datasets and cross-dataset settings. All results are directly reported from the original publications, ensuring a faithful reproduction of published claims. Metrics are presented as percentages (%). Abbreviations: CV = cross-validation; TL = transfer learning; CL = contrastive learning; DTW = dynamic time warping; AE = autoencoder; HN = hierarchical normalization. The proposed method (“Our”) demonstrates consistently competitive or superior performance, particularly in challenging cross-dataset generalization scenarios.

Setting	$\lambda_{\text{vq}}$	$\lambda_{\text{cong}}$	$\lambda_{\text{conl}}$	Chapman AUC	Chapman ACC	CPSC2018 AUC	CPSC2018 ACC
Default	0.2	0.5	0.1	99.98	99.66	99.91	99.37
Weaker global	0.2	0.1	0.1	99.74	98.83	99.48	98.26
Stronger global	0.2	1.0	0.1	99.80	99.03	99.62	98.63
Excessive global	0.2	3.0	0.1	99.22	97.59	98.51	96.54
Weaker local	0.2	0.5	0.02	99.66	98.54	99.35	98.07
Stronger local	0.2	0.5	0.20	99.70	98.75	99.54	98.42

Appendix

Table A.2: Sensitivity analysis of model performance to auxiliary loss weights, with

\lambda_{\text{cls}}=1

and

\lambda_{\text{rec}}=1

fixed. Reported metrics are AUC and accuracy (ACC) on Chapman and CPSC2018 datasets.

Figure A.3: Grad-CAM visualizations validate the functional specialization of our dual-encoder architecture. Color intensity reflects gradient-based importance: warm colors (red/yellow) indicate regions most influential to each encoder’s output, while cool colors (blue) denote low attention. (a) The rhythm encoder—using small kernels—consistently attends to the R-wave, its interval regularity directly underpins rhythm assessment and atrial fibrillation detection. (b) The morphology encoder—employing larger kernels—focuses on the P-wave region, which carries structural information about atrial depolarization: organized P-waves signify sinus rhythm, whereas their absence or disorganization is characteristic of atrial fibrillation. This clear dissociation confirms our design rationale: distinct receptive fields effectively decouple temporal dynamics (rhythm) from waveform structure (morphology).

Atrial Fibrillation Detection

See all 10 papers

Aug 4, 2026

FOUND-AF: Benchmarking ECG Foundation Models for Atrial Fibrillation Detection

Atrial fibrillation (AF) is the most common sustained cardiac arrhythmia and is associated with increased risks of stroke, heart failure, and mortality. Recent ECG foundation models offer transferable representations for automated AF detection. However, their relative effectiveness remains unclear because existing studies use different datasets, preprocessing procedures, classifiers, and validation protocols. This study presents FOUND-AF, a unified, leakage-controlled, and deployment-oriented benchmarking framework that evaluates the quality of pretrained ECG representations under identical experimental conditions. Nine publicly available foundation models from five families, including HuBERT-ECG, CLEF, ST-MEM, ECG-JEPA, and ECGFounder, were evaluated across four heterogeneous ECG datasets, namely AFDB, CinC2017, CPSC2021, and LTAFDB. All models were used as frozen feature extractors with standardized preprocessing, model-native resampling, a fixed XGBoost classifier, and recording-level grouped cross-validation. The evaluation included classification metrics, receiver operating characteristic analysis, paired recording-level bootstrap comparisons with Holm correction, embedding-space visualization, and computational efficiency profiling. The ECGFounder model consistently achieved the strongest overall performance across datasets while offering a favorable trade-off between accuracy, model size, inference time, and memory usage. FOUND-AF therefore provides a reproducible framework for selecting ECG foundation models and demonstrates that compact, clinically pretrained encoders can support robust and computationally efficient AF detection across heterogeneous acquisition settings.

Amirhossein Taleshinosrati, Yangyang Wang, Atitaya Phoemsuk +5

SDU Health Informatics and Technology University of Southern Denmark Odense, 5230, Denmark · Department of Computer Science University of Helsinki Helsinki, 00100, Finland · School of Computer Science and Electronic Engineering University of Essex Colchester, CO4 3SQ, UK +1

Atrial Fibrillation Detection Electrocardiogram Foundation Models Electrocardiogram Classification

Ask PDF

Dec 19, 2025

A Dataset and Benchmarks for Atrial Fibrillation Detection from Electrocardiograms of Intensive Care Unit Patients

Objective: Atrial fibrillation (AF) is the most common cardiac arrhythmia experienced by intensive care unit (ICU) patients and can cause adverse health effects. In this study, we publish a labelled ICU dataset and benchmarks for AF detection. Methods: We compared machine learning models across three data-driven artificial intelligence (AI) approaches: feature-based classifiers, deep learning (DL), and ECG foundation models (FMs). This comparison addresses a critical gap in the literature and aims to pinpoint which AI approach is best for accurate AF detection. Electrocardiograms (ECGs) from a Canadian ICU and the 2021 PhysioNet/Computing in Cardiology Challenge were used to conduct the experiments. Multiple training configurations were tested, ranging from zero-shot inference to transfer learning. Results: On average and across both datasets, ECG FMs performed best, followed by DL, then feature-based classifiers. The model that achieved the top F1 score on our ICU test set was ECG-FM through a transfer learning strategy (F1=0.89). Conclusion: This study demonstrates promising potential for using AI to build an automatic patient monitoring system. Significance: By publishing our labelled ICU dataset (LinkToBeAdded) and performance benchmarks, this work enables the research community to continue advancing the state-of-the-art in AF detection in the ICU. https://physionet.org/content/kingston-icu-af/

Sarah Nassar, Nooshin Maghsoodi, Sophia Mannina +7

Department of Electrical and Computer Engineering at Queen’s University, Kingston, ON, Canada · School of Computing at Queen’s University, Kingston, ON, Canada · Smith School of Business at Queen’s University, Kingston, ON, Canada +5

Atrial Fibrillation Detection Electrocardiogram Foundation Models Electrocardiogram Classification Time-Series Foundation Models

Ask PDF

May 23, 2026

HeartBeatAI: An Interpretable and Robust Deep Learning Framework for Multi-Label ECG Arrhythmia Detection

While Deep Learning (DL) enhances automated electrocardiogram (ECG) analysis, clinical deployment is hindered by class imbalance and the generalization gap. This paper presents HeartBeatAI, a deep learning framework combining domain generalization, multi-scale feature aggregation, and clinical explainability for robust 12-lead ECG classification. Moving beyond image-based paradigms, HeartBeatAI integrates a Squeeze-and-Excitation (SE) ResNet to isolate diagnostic leads alongside a Multi-Layer Concentration Pipeline to capture macro-rhythm and micro-morphological anomalies. To mitigate domain shift, the framework employs MixStyle regularization and Label Smoothing. Rigorous benchmarking across four large-scale datasets using intra-source and Leave-One-Domain-Out (LODO) protocols demonstrates high performance (98% Macro F1-score) under intra-source conditions. However, LODO evaluations reveal significant degradation in detecting rare anomalies, highlighting a persistent challenge in cross-institutional deployment.

Shubham Gupta, Nikhil Panwar, Partha Pratim Roy

Department of Computer Science and Engineering, Indian Institute of Technology (ISM) Dhanbad, Dhanbad, Jharkhand, India. · Department of Computer Science and Engineering, Indian Institute of Technology Roorkee, Roorkee, Uttarakhand, India.

Electrocardiography Domain Generalization Electrocardiogram Classification

Ask PDF

Abstract

Atrial Fibrillation Detection

Explore similar work

FOUND-AF: Benchmarking ECG Foundation Models for Atrial Fibrillation Detection

A Dataset and Benchmarks for Atrial Fibrillation Detection from Electrocardiograms of Intensive Care Unit Patients

HeartBeatAI: An Interpretable and Robust Deep Learning Framework for Multi-Label ECG Arrhythmia Detection