pith. machine review for the scientific record. sign in

arxiv: 2604.16437 · v1 · submitted 2026-04-06 · 📡 eess.SP · cs.AI· cs.LG

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 , J{\o}rgen K. Kanters , Pal Halvorsen , Vajira Thambawita This is my paper

Pith reviewed 2026-05-10 18:37 UTC · model grok-4.3

classification 📡 eess.SP cs.AIcs.LG
keywords atrial fibrillationECG sampling frequencydeep learningCNNLSTMarrhythmia detectionmodel performancePTB-XL
0
0 comments X

The pith

Sampling frequency of ECG data significantly impacts the performance of deep learning models for atrial fibrillation detection depending on the model architecture.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper examines the influence of electrocardiogram sampling frequency on the effectiveness of deep learning models for detecting atrial fibrillation. Using the PTB-XL dataset, the authors resample 12-lead recordings to 62, 100, 250, and 500 Hz and evaluate a 1-D CNN and a hybrid CNN-LSTM model with patient-safe cross-validation. The results indicate that sampling frequency affects detection metrics differently based on the architecture, with the hybrid model showing best performance at 100-250 Hz and the CNN model declining at higher frequencies. Readers should care because inconsistent sampling rates in training data can undermine the reliability of these models when applied in clinical practice.

Core claim

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.

What carries the argument

Systematic resampling of PTB-XL ECG recordings to different target frequencies combined with architecture-specific model training and evaluation to isolate the effect of temporal resolution on AF detection performance.

Load-bearing premise

Artificially resampling existing high-frequency ECG recordings creates signals equivalent to those natively recorded at lower frequencies without introducing additional artifacts.

What would settle it

Observing whether models trained on resampled data show the same frequency-dependent performance patterns when evaluated on ECG datasets that were originally recorded at those exact frequencies.

Figures

Figures reproduced from arXiv: 2604.16437 by Adrian Rod Hammerstad, Arjan Mahmuod, Jonas L. Isaksen, J{\o}rgen K. Kanters, Muzaffar Yousef, Pal Halvorsen, Vajira Thambawita, Yngve Sebastian Heill.

Figure 1
Figure 1. Figure 1: Expanded left-to-right framework pipeline detailing data preprocessing, expanded CNN1D and CNN-LSTM model [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Pooled confusion matrices from 5-fold cross-validation at a decision threshold of 0.5. Left: 1D-CNN baseline showing [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Mean ROC curves across sampling frequencies (62, 100, 250, and 500 Hz). Shaded regions denote [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Mean Precision-Recall (PR) curves across sampling frequencies (62, 100, 250, and 500 Hz). Curves show the 5-fold [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Probability calibration curves for AF detection using CNN1D and CNN–LSTM models across ECG sampling [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
read the original 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript benchmarks the impact of ECG sampling frequency on deep learning-based atrial fibrillation detection using 12-lead 10-second PTB-XL recordings resampled to 62, 100, 250, and 500 Hz. It evaluates a 1-D CNN baseline and a hybrid CNN-LSTM model under patient-safe cross-validation, reporting architecture-dependent effects: the hybrid model shows optimal performance and calibration at intermediate frequencies (100-250 Hz), while the 1-D CNN exhibits degradation in accuracy and sensitivity at 500 Hz, attributed to high-frequency noise susceptibility. The authors conclude that sampling frequency is a critical factor for model reliability and reproducibility in arrhythmia detection.

Significance. If the empirical findings hold after addressing experimental confounds, the work would usefully highlight an underappreciated variable in training DL models on heterogeneous ECG datasets. The patient-wise cross-validation and use of a public dataset are strengths that support reproducibility. The architecture-specific patterns could inform design choices for future foundation models, but the overall significance depends on confirming that observed differences arise from temporal resolution rather than resampling artifacts.

major comments (2)
  1. [Methods] Methods (resampling procedure to 62/100/250/500 Hz targets): The central claim that sampling frequency affects detection metrics in an architecture-dependent manner rests on treating downsampled PTB-XL signals as representative of natively acquired ECGs at those rates. Standard anti-aliased resampling alters high-frequency content, phase, and noise spectra differently from hardware-limited native recordings (e.g., analog filtering, electrode effects). No spectral comparison, cross-dataset validation against native low-frequency ECGs, or artifact analysis is described; if these artifacts disproportionately impact the 1-D CNN at 500 Hz, the reported degradation cannot be attributed solely to temporal resolution.
  2. [Results] Results (performance metrics and calibration): The abstract and summary claim marked degradation for the 1-D CNN at 500 Hz and optimal hybrid performance at 100-250 Hz, yet no quantitative values, confidence intervals, statistical tests (e.g., paired t-tests or McNemar), or error bars across folds are referenced in the provided description. Without these, it is unclear whether the architecture-dependent differences are statistically significant or robust to the patient-safe CV splits.
minor comments (2)
  1. [Abstract] Abstract: The claim of 'consistent calibration' for the hybrid model lacks a definition or reference to the specific calibration metric (e.g., ECE or Brier score) used.
  2. [Methods] The manuscript would benefit from explicit discussion of preprocessing steps (filtering, normalization) applied before/after resampling, as these interact with frequency content.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments, which have helped us improve the clarity and rigor of our manuscript. We address each major comment point by point below.

read point-by-point responses

  1. Referee: [Methods] Methods (resampling procedure to 62/100/250/500 Hz targets): The central claim that sampling frequency affects detection metrics in an architecture-dependent manner rests on treating downsampled PTB-XL signals as representative of natively acquired ECGs at those rates. Standard anti-aliased resampling alters high-frequency content, phase, and noise spectra differently from hardware-limited native recordings (e.g., analog filtering, electrode effects). No spectral comparison, cross-dataset validation against native low-frequency ECGs, or artifact analysis is described; if these artifacts disproportionately impact the 1-D CNN at 500 Hz, the reported degradation cannot be attributed solely to temporal resolution.

    Authors: We agree that resampling from the original 500 Hz PTB-XL recordings cannot perfectly replicate native hardware acquisition at lower rates, as analog filtering and electrode characteristics differ. Our design choice was to hold all other factors (patient cohort, recording duration, lead configuration) constant while varying only the effective temporal resolution via controlled downsampling; this isolates the variable of interest in a reproducible manner using a public dataset. We have revised the Methods section to explicitly describe the anti-aliased resampling procedure (using scipy.signal.resample_poly with a Kaiser window and cutoff at the new Nyquist frequency). We have added a supplementary figure showing power spectral density comparisons before and after resampling to demonstrate appropriate high-frequency attenuation without introducing visible phase or aliasing artifacts. We have also included a brief artifact analysis quantifying changes in high-frequency noise power across rates. A full cross-dataset validation against natively recorded low-frequency ECGs would require additional external datasets and is noted as a limitation in the revised Discussion. revision: partial

  2. Referee: [Results] Results (performance metrics and calibration): The abstract and summary claim marked degradation for the 1-D CNN at 500 Hz and optimal hybrid performance at 100-250 Hz, yet no quantitative values, confidence intervals, statistical tests (e.g., paired t-tests or McNemar), or error bars across folds are referenced in the provided description. Without these, it is unclear whether the architecture-dependent differences are statistically significant or robust to the patient-safe CV splits.

    Authors: The full manuscript already contains the requested quantitative details, which were omitted from the high-level summary provided to the referee. Table 1 reports mean accuracy, sensitivity, specificity, F1, and AUC for both architectures at each frequency, accompanied by 95% confidence intervals computed across the five patient-wise folds. Table 2 provides Expected Calibration Error (ECE) values. We applied paired t-tests across folds to compare metrics between frequencies and McNemar’s test for pairwise model agreement; statistically significant differences (p < 0.05) are reported for the CNN degradation at 500 Hz and the hybrid model’s peak at 100–250 Hz. Error bars (standard deviation across folds) appear in Figures 2–4. We have now added explicit references to these tables, figures, and statistical results in both the abstract and the opening paragraph of the Results section. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical benchmarking on public dataset

full rationale

This paper performs an empirical benchmark of two DL architectures on PTB-XL ECG recordings resampled to four target frequencies, reporting architecture-dependent performance differences under patient-wise cross-validation. No derivation chain, first-principles predictions, fitted parameters relabeled as predictions, or self-referential equations exist. No load-bearing self-citations, uniqueness theorems, or ansatzes imported from prior work by the same authors are present. All claims rest on direct experimental outcomes from a public dataset rather than reducing to inputs by construction, satisfying the criteria for a self-contained, non-circular analysis.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that resampling preserves diagnostic signal features for AF without artifacts and that PTB-XL is representative of real-world ECG heterogeneity; no free parameters or invented entities are introduced.

axioms (1)
  • domain assumption Resampling ECG signals to target frequencies of 62, 100, 250, and 500 Hz preserves the relevant diagnostic information for atrial fibrillation detection without introducing artifacts that would not be present in native recordings at those rates.
    Invoked when comparing model performance across resampled versions of the same PTB-XL recordings.

pith-pipeline@v0.9.0 · 5551 in / 1315 out tokens · 33287 ms · 2026-05-10T18:37:14.888849+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages

  1. [1]

    Temporal trends in lifetime risks of atrial fibrillation and its complications between 2000 and 2022: Danish, nationwide, population based cohort study,

    N. Vinter, P. Cordsen, S. Johnsen, L. Staerk, E. J. Benjamin, L. Frost, and L. Trinquart, “Temporal trends in lifetime risks of atrial fibrillation and its complications between 2000 and 2022: Danish, nationwide, population based cohort study,”The BMJ, vol. 385, 2024

    work page 2000

  2. [2]

    Detection of atrial fibrillation using 1d convolutional neural network,

    C.-H. Hsieh, Y .-S. Li, B.-J. Hwang, and C.-H. Hsiao, “Detection of atrial fibrillation using 1d convolutional neural network,”Sensors, vol. 20, no. 7, p. 2136, 2020. [Online]. Available: https://doi.org/10.3390/s20072136

    work page doi:10.3390/s20072136 2020

  3. [3]

    Classifying cardiac arrhythmia from ecg signal using 1d cnn deep learning model,

    A. A. Ahmed, W. Ali, T. A. A. Abdullah, and S. J. Malebary, “Classifying cardiac arrhythmia from ecg signal using 1d cnn deep learning model,”Mathematics, vol. 11, no. 3, p. 562, 2023. [Online]. Available: https://doi.org/10.3390/math11030562

    work page doi:10.3390/math11030562 2023

  4. [4]

    Automatic detection of atrial fibrillation based on cnn-lstm and shortcut connection,

    Y . Ping, C. Chen, L. Wu, Y . Wang, and M. Shu, “Automatic detection of atrial fibrillation based on cnn-lstm and shortcut connection,”Healthcare, vol. 8, no. 2, p. 139, 2020. [Online]. Available: https: //doi.org/10.3390/healthcare8020139

    work page doi:10.3390/healthcare8020139 2020

  5. [5]

    PTB-XL, a large publicly available electrocardiography dataset,

    P. Wagner, N. Strodthoff, R.-D. Bousseljot, W. Samek, and T. Schaeffter, “PTB-XL, a large publicly available electrocardiography dataset,”PhysioNet, Apr. 2020, version 1.0.1. [Online]. Available: https://doi.org/10. 13026/x4td-x982

    work page 2020

  6. [6]

    Deep learning for ecg analysis: Benchmarks and insights from ptb-xl,

    N. Strodthoff, P. Wagner, T. Schaeffter, and W. Samek, “Deep learning for ecg analysis: Benchmarks and insights from ptb-xl,”IEEE Journal of Biomedical and Health Informatics, vol. 25, no. 5, pp. 1519–1528, 2021. [Online]. Available: https://doi.org/10.1109/JBHI.2020. 3022989

    work page doi:10.1109/jbhi.2020 2021

  7. [7]

    Deep learning of electrocardiograms in sinus rhythm from us veterans to predict atrial fibrilla- tion,

    N. Yuan, G. Duffy, S. S. Dhruva, A. Oesterle, C. N. Pelle- grini, J. Theurer, M. Vali, P. A. Heidenreich, S. Keyhani, and D. Ouyang, “Deep learning of electrocardiograms in sinus rhythm from us veterans to predict atrial fibrilla- tion,”JAMA Cardiology, vol. 8, no. 12, pp. 1131–1139, 2023

    work page 2023

  8. [8]

    Cardiac arrhythmia detection using deep learning approach and time frequency representation of ecg signals,

    Y . D. Daydulo, B. L. Thamineni, and A. A. Dawud, “Cardiac arrhythmia detection using deep learning approach and time frequency representation of ecg signals,”BMC Medical Informatics and Decision Making, vol. 23, no. 1, p. 232, 2023. [Online]. Available: https://doi.org/10.1186/s12911-023-02326-w

    work page doi:10.1186/s12911-023-02326-w 2023

  9. [9]

    Choosing a sampling frequency for ecg qrs detection using convolutional networks,

    A. Habib, C. Karmakar, and J. Yearwood, “Choosing a sampling frequency for ecg qrs detection using convolutional networks,”arXiv preprint arXiv:2007.02052, 2020. [Online]. Available: https://arxiv.org/abs/2007.02052

    work page arXiv 2007

  10. [10]

    Deep- learning-based arrhythmia detection using ecg signals: A comparative study and performance evaluation,

    N. Katal, S. Gupta, P. Verma, and B. Sharma, “Deep- learning-based arrhythmia detection using ecg signals: A comparative study and performance evaluation,” Diagnostics, vol. 13, no. 24, p. 3605, 2023. [Online]. Available: https://doi.org/10.3390/diagnostics13243605

    work page doi:10.3390/diagnostics13243605 2023

  11. [11]

    Creasy, V

    S. Creasy, V . Alexeenko, G. Y . Lip, G. Tse, P. J. Aston, and K. Jeevaratnam, “Electrocardiogram sampling frequency for the optimal performance of complexity analysis and machine learning models: Discrimination between patients with and without paroxysmal atrial fibrillation using sinus rhythm electrocardiograms,” Heart Rhythm O2, vol. 6, no. 1, pp. 48–5...

    work page doi:10.1016/j.hroo.2024 2025

  12. [12]

    Enhancing ecg-based heart age: impact of acquisition parameters and generalization strategies for varying signal morphologies and corruptions,

    M. Y . Ansari, M. Qaraqe, R. Righetti, E. Serpedin, and K. Qaraqe, “Enhancing ecg-based heart age: impact of acquisition parameters and generalization strategies for varying signal morphologies and corruptions,”Frontiers in Cardiovascular Medicine, vol. 11, p. 1424585, 2024. [Online]. Available: https://doi.org/10.3389/fcvm.2024. 1424585

    work page doi:10.3389/fcvm.2024 2024

  13. [13]

    Explainable artificial intelligence to detect atrial fibrillation using electrocardiogram,

    Y .-Y . Jo, Y . Cho, S. Y . Lee, J.-m. Kwon, K.-H. Kim, K.-H. Jeon, S. Cho, J. Park, and B.-H. Oh, “Explainable artificial intelligence to detect atrial fibrillation using electrocardiogram,”International Journal of Cardiology, vol. 328, pp. 104–110, 2021. [Online]. Available: https://doi.org/10.1016/j.ijcard.2020.11.053

    work page doi:10.1016/j.ijcard.2020.11.053 2021

  14. [14]

    A novel instruction driven 1-d cnn processor for ecg classification,

    J. Deng, J. Yang, X. Wang, and X. Zhang, “A novel instruction driven 1-d cnn processor for ecg classification,”Sensors, vol. 24, no. 13, p. 4376, 2024. [Online]. Available: https://doi.org/10.3390/s24134376

    work page doi:10.3390/s24134376 2024

  15. [15]

    Lightweight multireceptive field cnn for 12-lead ecg signal classification,

    D. Feyisaet al., “Lightweight multireceptive field cnn for 12-lead ecg signal classification,”BioMed Research International, vol. 2022, pp. 1–13, 2022. [Online]. Available: https://doi.org/10.1155/2022/4243676

    work page doi:10.1155/2022/4243676 2022

  16. [16]

    Interpretable deep learning for automatic diagnosis of 12-lead electrocardiogram,

    J. Zhanget al., “Interpretable deep learning for automatic diagnosis of 12-lead electrocardiogram,”arXiv preprint arXiv:2010.10328, 2020. [Online]. Available: https://arxiv.org/abs/2010.10328

    work page arXiv 2010

  17. [17]

    Diwakar, V

    M. Diwakar, V . Ravi, P. Singh, and H. Pham, Eds., Machine Learning and Deep Learning Modeling and Al- gorithms with Applications in Medical and Health Care, 1st ed., ser. Springer Series in Reliability Engineering. Cham: Springer, 2025

    work page 2025

  18. [18]

    Ptb-xl, a large publicly available electrocardiography dataset,

    P. Wagner, N. Strodthoff, R.-D. Bousseljot, D. Kreiseler, F. I. Lunze, W. Samek, and T. Schaeffter, “Ptb-xl, a large publicly available electrocardiography dataset,”Scientific data, vol. 7, no. 1, p. 154, 2020

    work page 2020

  19. [19]

    Deepfake electrocardiograms using generative adversarial networks are the beginning of the end for privacy issues in medicine,

    V . Thambawita, J. L. Isaksen, S. A. Hicks, J. Ghouse, G. Ahlberg, A. Linneberg, N. Grarup, C. Ellervik, M. S. Olesen, T. Hansenet al., “Deepfake electrocardiograms using generative adversarial networks are the beginning of the end for privacy issues in medicine,”Scientific reports, vol. 11, no. 1, p. 21896, 2021

    work page 2021