Sampling Matters: The Effect of ECG Frequency on Deep Learning-Based Atrial Fibrillation Detection

Arjan Mahmuod; Adrian Rod Hammerstad; Muzaffar Yousef; Yngve Sebastian Heill; Jonas L. Isaksen; Jørgen K. Kanters; Pal Halvorsen; Vajira Thambawita

eess.SPApr 6, 2026

Sampling Matters: The Effect of ECG Frequency on Deep Learning-Based Atrial Fibrillation Detection

Authors: Arjan Mahmuod, Adrian Rod Hammerstad, Muzaffar Yousef, Yngve Sebastian Heill, Jonas L. Isaksen, Jørgen K. Kanters, Pal Halvorsen, Vajira Thambawita

Organizations: Oslo Metropolitan University, Norway · University of Copenhagen, Denmark · SimulaMet, Norway

Abstract

Deep learning models for atrial fibrillation (AF) detection are increasingly trained on heterogeneous electrocardiogram (ECG) datasets with varying sampling frequencies, yet the specific consequences of these discrepancies on model performance, calibration, and robustness remain insufficiently characterized. To address this, we conducted a systematic benchmark using 12-lead, 10-second recordings from the PTB-XL dataset, resampled to target frequencies of 62, 100, 250, and 500 Hz, to evaluate a standard 1-D Convolutional Neural Network (CNN) and a hybrid CNN-Long Short-Term Memory (LSTM) architecture under a rigorous patient-safe cross-validation framework. Our analysis reveals that sampling frequency significantly impacts detection metrics in an architecture-dependent manner; the hybrid CNN-LSTM model demonstrated optimal performance and consistent calibration at intermediate frequencies (100-250 Hz), whereas the 1-D CNN baseline exhibited marked degradation in accuracy and sensitivity at 500 Hz, suggesting increased susceptibility to high-frequency noise. We conclude that ECG sampling frequency is a critical, underappreciated factor in arrhythmia detection, and future foundation models must explicitly control for temporal resolution to ensure clinical reliability and reproducibility.

Figures & tables

Study	12L	AF	Freq:	Calib:	Repro:
Daydulo et al. [4]	✗	★	✓	★	✗
Habib et al. [8]	★	✗	✓	✗	✗
Katal et al. [11]	✗	✓	✗	★	✗
Creasy et al. [3]	✓	✓	✓	✗	★
Ansari et al. [2]	✓	✗	✓	✗	✓
Yuan et al. [18]	✓	✓	✗	✓	✗

TABLE I : Comparison of Existing AF Detection Studies with our study. ( 12L =12-Lead ECG, AF =AF Detection, Freq: =Frequency Analysis, Calid: =Calibration Analysis, Repro: =Reproducible Code, ✓ = explicitly addressed, ✗ = not addressed, ★ = partially addressed)

Panel of Figure 2 — Fig. 1 : Expanded left-to-right framework pipeline detailing data preprocessing, expanded CNN1D and CNN-LSTM model architectures, and test evaluation phases. Row boundaries and inter-row spacing are completely uniform.

Dataset	Model	Frequency	Accuracy	F1	Precision	Sensitivity	Specificity	MCC	AUROC	ECE $\downarrow$
Validation	CNN–LSTM	62 Hz	$0.9844\pm 0.0074$	$0.9844\pm 0.0075$	$0.9873\pm 0.0078$	$0.9816\pm 0.0156$	$0.9873\pm 0.0076$	$0.9691\pm 0.0149$	$0.9952\pm 0.0040$	$0.021\pm 0.002$
	CNN–LSTM	100 Hz	$0.9821\pm 0.0087$	$0.9820\pm 0.0089$	$0.9855\pm 0.0103$	$0.9788\pm 0.0165$	$0.9854\pm 0.0102$	$0.9644\pm 0.0173$	$0.9937\pm 0.0043$	$0.028\pm 0.012$
	CNN–LSTM	250 Hz	$\mathbf{0.9878\pm 0.0047}$	$\mathbf{0.9877\pm 0.0048}$	$\mathbf{0.9922\pm 0.0066}$	$0.9835\pm 0.0135$	$\mathbf{0.9922\pm 0.0065}$	$\mathbf{0.9758\pm 0.0095}$	$0.9941\pm 0.0060$	$0.020\pm 0.006$
	CNN–LSTM	500 Hz	$0.9849\pm 0.0090$	$0.9847\pm 0.0091$	$0.9911\pm 0.0038$	$0.9786\pm 0.0175$	$0.9912\pm 0.0037$	$0.9701\pm 0.0180$	$0.9963\pm 0.0034$	$0.026\pm 0.013$
	CNN1D	62 Hz	$0.9859\pm 0.0052$	$0.9858\pm 0.0053$	$0.9893\pm 0.0032$	$0.9825\pm 0.0100$	$0.9893\pm 0.0033$	$0.9719\pm 0.0106$	$0.9956\pm 0.0041$	$0.020\pm 0.005$
	CNN1D	100 Hz	$0.9874\pm 0.0079$	$0.9873\pm 0.0080$	$0.9893\pm 0.0058$	$\mathbf{0.9854\pm 0.0150}$	$0.9893\pm 0.0057$	$0.9748\pm 0.0159$	$\mathbf{0.9966\pm 0.0037}$	$\mathbf{0.018\pm 0.006}$

TABLE II : 5-fold cross-validation and final imbalanced test performance across ECG sampling frequencies. Validation metrics are reported as mean

\pm

standard deviation across folds. ECE

\downarrow

indicates lower is better. Best overall values per dataset (Validation and Test) are highlighted in bold . Best values for each individual model within a dataset are highlighted in red .

Panel of Figure 4 — Fig. 2 : Pooled confusion matrices from 5-fold cross-validation at a decision threshold of 0.5. Left: 1D-CNN baseline showing degradation at high frequency. Right: Hybrid CNN-LSTM showing consistent performance across all frequencies. Note the variation in error counts at 500 Hz demonstrates stable error rates across the frequency spectrum.

Panel of Figure 5 — Fig. 3 : Mean ROC curves across sampling frequencies (62, 100, 250, and 500 Hz). Shaded regions denote $\pm 1$ standard deviation (x-axis: false positive rate, y-axis: true positive rate).

Panel of Figure 6 — Fig. 4 : Mean Precision-Recall (PR) curves across sampling frequencies (62, 100, 250, and 500 Hz). Curves show the 5-fold cross-validation mean; shaded regions denote $\pm 1$ standard deviation (x-axis: Precision, y-axis: Recall).

Fig. 5 : Probability calibration curves for AF detection using CNN1D and CNN–LSTM models across ECG sampling frequencies. The dashed diagonal denotes perfect calibration; curves closer to it indicate better probabilistic calibration.

Atrial Fibrillation Detection

See all 10 papers

FOUND-AF: Benchmarking ECG Foundation Models for Atrial Fibrillation Detection
Aug 4, 2026Amirhossein Taleshinosrati, Yangyang Wang, Atitaya Phoemsuk +5Atrial Fibrillation Detection Electrocardiogram Foundation Models
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 with Arbitrary Leads via a Codebook-Based Reconstruction-Classification Framework
Aug 19, 2026Hongtao Li, Jia Wei, Guoyao Li +6Atrial Fibrillation Detection Electrocardiogram Classification
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 +1
A Dataset and Benchmarks for Atrial Fibrillation Detection from Electrocardiograms of Intensive Care Unit Patients
Dec 19, 2025Sarah Nassar, Nooshin Maghsoodi, Sophia Mannina +7Atrial Fibrillation Detection Electrocardiogram Foundation Models
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

Aug 4, 2026J/K move · Enter open · S save

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

Amirhossein Taleshinosrati, Yangyang Wang, Atitaya Phoemsuk, Vahid Abolghasemi+4

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

Open paper AskPDF arXiv

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.

Abstract

Atrial Fibrillation Detection

Explore similar work

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

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

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