Database Open Access

Noise Holter Database: a multi-device, multi-arrhythmia real-world clinical ECG signal quality benchmark

,  ,  ,  , 

Published Nov. 17, 2026 · Version 1.0.0
When using this resource, please cite:

Fleury, Q., Dubois, R., Christophle-Boulard, S., Extramiana, F., & Maison-Blanche, P. (2026). Noise Holter Database: a multi-device, multi-arrhythmia real-world clinical ECG signal quality benchmark (version 1.0.0). PhysioNet. RRID:SCR_007345. https://doi.org/10.13026/8kac-hn03

Please include the standard citation for PhysioNet:

Pollard, T., Moody, B. E., Lehman, L., Gow, B., Fernandes, C., Xie, C., Johnson, A., Mark, R. G., & Heldt, T. (2026). PhysioNet as a global platform for biomedical research. Nature Health. https://doi.org/10.1038/s44360-026-00096-z. Available from: https://pmc.ncbi.nlm.nih.gov/articles/PMC13641116/

Abstract

The Noise Holter Database is a publicly available, multi-arrhythmia Holter ECG dataset comprising 258 two-lead recordings from distinct subjects. The primary objective of making this resource available is to facilitate the development of software for filtering out noisy segments in ambulatory recordings. The dataset includes a wide spectrum of spontaneous cardiac arrhythmias—including atrial fibrillation, ventricular tachycardia, AV block, and supraventricular tachycardia—captured under real-world ambulatory conditions with authentic environmental noise. Encompassing approximately 35 hours of data, the database features 6,255 ECG segments. These segments include expert-assigned multi-label signal quality annotations classified into three functional levels: Clean (43.7%), N1 where QRS is identifiable (36.0%), and N2 where QRS is unidentifiable (20.3%). Inter-rater agreement, assessed on a comprehensive validation subset of 2,166 events, yielded a Cohen's kappa of 0.83, demonstrating the robustness of the labeling process. This dataset effectively addresses the critical absence of publicly available benchmarks with authentic, naturally occurring ambulatory noise annotations.


Background

Cardiovascular diseases remain a leading cause of global mortality, and long-term ECG recordings are the backbone of clinical management for cardiac arrhythmias. However, the reliability of automated analysis software is directly conditioned by signal quality; noise present in ambulatory recordings significantly degrades the performance of both rule-based and deep learning algorithms [1]. Unlike resting ECGs, Holter recordings are susceptible to severe artifacts like skeletal muscle activity and electrode-skin motion. While several public ECG databases exist, datasets specifically focusing on real-world clinical noise are scarce. Existing resources, such as the MIT-BIH Noise Stress Test Database [2], frequently rely on artificially superimposed noise rather than authentic, naturally occurring ambulatory artifacts. Other publicly available ECG resources generally lack real-world noise annotations across diverse arrhythmias, including the BUT QDB [3], the MIT-BIH Arrhythmia Database [4] [5], IRIDIA-AF [6], SHDB-AF [7], and LTAFDB [8]. The motivation for sharing this resource is to establish a benchmark that relies on unsimulated data and defines the practical noise thresholds at which clinical experts can confidently identify ECG waveforms.


Methods

The source Holter ECG recordings were acquired from patients referred to the Cardiac Arrhythmias Unit at Centre Médico-Chirurgical Ambroise Paré (Neuilly-sur-Seine, France) as part of routine clinical care. Holter monitoring was performed on physician prescription according to clinical indications, independently of the present research project. The recordings were routinely archived in the clinical Holter database and subsequently anonymized before being made available for construction of the Noise Holter Database. The ECG signals were sampled at 200Hz with a 15-bit resolution for SpiderView, and at 256Hz for SpiderFlash, SpiderSAS (both 15-bit), and DMS myPatch (12-bit) devices. Prior to manual noise labeling, recordings were annotated for heartbeat detection and arrhythmia occurrences using SyneScope automated software. An expert cardiologist then manually adjusted QRS annotations and validated the arrhythmias.

Signal quality within the database is classified into three distinct functional levels to evaluate the feasibility of automated and clinical analysis, for each ECG lead [3]:

  • Clean: Represents an optimal signal where atrial and ventricular electrical activities (such as normal waveforms or sustained arrhythmia patterns) are clearly readable without impact on the signal interpretation.
  • N1: Represents an intermediate noise level where the signal is corrupted by artifacts, but the QRS complexes can still be reliably identified. This class is clinically relevant as it represents the threshold at which rhythm analysis remains feasible.
  • N2: Designates severe noise where atrial and ventricular electrical activities are completely obscured and no longer clearly recognizable. Segments in this class are unsuitable for automated or manual ECG editing.

This three-class delineation is motivated by practical clinical decision-making: the ability to detect a QRS complex is the minimum requirement for cardiac rhythm analysis, while P-wave visibility determines the feasibility of enhanced arrhythmia characterization. The N1/N2 boundary therefore reflects the clinical threshold for automated algorithm reliability.

The database construction utilized two complementary sampling protocols:

  • Targeted manual delineation: Applied to a primary group of 157 recordings, where an expert manually selected variable-length segments covering specific noise events.
  • Stratified sampling: Applied to a secondary group of 101 recordings. A 1D-ResNet deep learning model pre-screened the continuous data to flag potential segments of interest, particularly complex edge cases where severe noise coincided with ectopic activity. Fixed 5-second segments were then extracted and visually inspected by an expert to establish the clinical ground truth.

To validate the reliability of the clinical labels, a secondary expert blindly reviewed a validation subset of 2,166 events. This independent evaluation yielded a Cohen's kappa coefficient of 0.83. Following this review, a consensus reconciliation process was conducted to resolve any disagreements and establish a single, unified ground truth annotation.


Data Description

Signal Quality Annotation

Combining both annotation protocols, the signal quality annotated dataset comprises 6,255 ECG segments (4,405 variable-length segments from the primary group and 1,850 fixed 5-second segments from the secondary group), representing approximately 35 hours of recording across all 258 patients.

Each extracted segment is characterized by a single global noise event, with the distribution of these events across quality classes and acquisition subsets detailed in Table 1.

Table 1. Distribution of signal quality annotations across quality classes (Clean, N1, N2) and acquisition subsets (Primary and Secondary groups). N, number of annotated events; Duration in seconds.
Quality Noise Type N Strips Total Duration (s) Min (s) Max (s)
Clean Primary group 1,768 68,188 6 178
Secondary group 968 4,840 5 5
Clean Total 2,736 73,028 5 178
N1 Primary group 1,544 23,853 4 175
Secondary group 706 3,530 5 5
N1 Total 2,250 27,383 4 175
N2 Primary group 1,093 24,963 4 349
Secondary group 176 880 5 5
N2 Total 1,269 25,843 4 349
Overall Total 6,255 126,254

Clinical event annotation

Furthermore, because the database utilizes a multi-label clinical annotation methodology, these 6,255 physical segments yield a richer subset of distinct clinical event annotations. The global distribution of these underlying pathological events, independent of their signal quality class, is presented in Table 2.

Table 2. Distribution of clinical event annotations. N, number of annotated events; Duration in seconds. Note: Due to the multi-label annotation, a single segment can appear multiple times in this table across different event categories.
Event Type N Strips Total Duration (s)
NSR 3,672 58,534
A. Fib. 379 14,578
A. Flutter 122 8,626
Atrial Tachycardia 14 70
AVB 368 10,640
Brady 210 8,733
MB 29 2,406
SV Couplet 108 3,286
SV Tachycardia 42 2,390
SVPB 852 17,063
TgSV 99 2,300
V Couplet 83 3,204
VPB 913 22,780
Overall Total 6,891 154,610

Data Format and File Structure

All ECG waveform data and expert annotations are provided as extracted segments in the standard PhysioNet/WFDB (Waveform Database) format. Prior to final segmentation, the continuous Holter recordings from the 258 subjects were systematically exported as discrete 24-hour (daily) records. All recordings were irreversibly anonymized and assigned new sequential identifiers. Each clinical segment comprises four interconnected files. The signal data is stored in a 16-bit binary file (.dat) containing the raw ECG waveform. Crucially, the number of leads included in each extracted segment depends directly on the annotation context: a segment contains two concurrent leads if both channels shared the exact same signal quality and clinical annotation, or a single lead if the specific annotated event applied to only one channel. This is accompanied by a standard ASCII header file (.hea) detailing critical metadata required for signal processing, including the 200 Hz sampling frequency, number of leads, ADC resolution, lead specifications, recording duration, and the specific Holter device type. The clinical ground truth is provided in a binary annotation file (.atr), where the temporal locations of individual beats, expressed as sample indices, are associated with standard WFDB annotation symbols ('N' for normal sinus beats, 'V' for premature ventricular contractions, 'S' for supraventricular premature beats, etc.). Additionally, a supplementary binary file (.qrs) contains exclusively the temporal indices of detected QRS complexes.

Signal Quality and Arrhythmia Label Encoding

Because the database consists of extracted segments rather than continuous Holter recordings, the global characteristics of each segment are embedded directly within the comment section of the .hea file. To enable seamless automated parsing and explicitly separate technical artifacts from clinical pathologies, these global annotations follow a strict, unified string convention: CLASS:<Quality> | NOISE:<Noise_Type> | EVENTS:<Pathologies>. Under this convention, a segment heavily affected by mid-frequency noise but capturing an underlying atrial fibrillation and a premature ventricular beat is encoded as CLASS: N1 | NOISE:Mid Frequency | EVENTS: A. Fib. | VPB. If a segment exhibits multiple simultaneous pathologies, they are automatically deduplicated, sorted alphabetically, and concatenated using a pipe character (“|”). While the .hea file provides the segment's global summary, the specific temporal context is preserved within the WFDB aux_note field of the .atr files.

The <Noise_Type> variable classifies the primary artifact source based on expert visual assessment. The possible values include Clean (artifact-free signal), Baseline Wander (low-frequency drifts), Electromagnetic Noise, High Frequency (typically muscle artifacts), Mid Frequency (mixed noise sources), Motion Artifact (abrupt baseline shifts due to movement), 50 Hz Power Line Interference, and Unknown (unidentifiable origin).


Usage Notes

The Noise Holter Database is intended for the development and evaluation of automated signal quality assessment algorithms for Holter ECG recordings. The three-class quality strip labelling (Clean, N1, N2) enables supervised training of classification models targeting real-world ambulatory noise levels.

The dataset is also suitable for multi-task learning approaches that jointly perform arrhythmia detection and signal quality assessment, given the concurrent availability of arrhythmia annotations and quality labels in the annotated subset.

Users should be aware of certain methodological constraints when using this database. First, the source recordings were acquired as part of routine clinical care at a single clinical center (Centre Médico-Chirurgical Ambroise Paré), which may limit how well the noise profiles generalize to other populations and clinical settings. Second, four commercial Holter devices were used (SpiderView, SpiderFlash, SpiderSAS, DMS myPatch), but most recordings were acquired with the SpiderView (75.9 % of segments). Device heterogeneity is therefore limited, and the noise characteristics mainly reflect this one device. Finally, all annotations are based on visual adjudication by expert cardiologists, which is the standard for clinical ECG interpretation. Signal quality levels (Clean, N1, N2) showed strong inter-rater agreement (Cohen's κ = 0.83). The <Noise_Type> labels, however, should be interpreted with more caution. Telling noise sources apart from morphology alone (for example mid-frequency vs. high-frequency noise, or motion artifact vs. baseline wander) is inherently less reliable than judging whether waveforms are readable, and these labels were not part of the inter-rater agreement assessment.


Release Notes

Version 1.0.0

Initial release


Ethics

All Holter ECG recordings included in this database were originally acquired at Centre Médico-Chirurgical Ambroise Paré (Neuilly-sur-Seine, France) as part of routine clinical care, on physician prescription and independently of the present research project. No additional examination, intervention, or ECG recording was performed for research purposes.

At the time of clinical care, patients were informed that their Holter recordings could subsequently be used in anonymized form for scientific purposes, and their oral agreement to such secondary use was obtained.

Recordings selected for the present database were extracted and irreversibly anonymized within the clinical environment before being made available for research. All patient identifiers and individual demographic and clinical information were removed. New sequential identifiers were assigned to the recordings, and no correspondence table, re-identification key, or other mechanism enabling linkage to the original patients was created or retained. The research team had access only to the anonymized ECG recordings and aggregate demographic statistics and had no access to the patients’ medical records or to individual demographic, clinical, or identifying information.


Acknowledgements

The authors acknowledge funding from the National Research Agency (ANR-10-IAHU-04 LIRYC).


Conflicts of Interest

The authors have no conflicts of interest to declare.


References

  1. Guerra RdT, Yamaguchi CK, Stefenon SF, Coelho LdS, Mariani VC (2025). "Deep Learning Approach for Automatic Heartbeat Classification". Sensors. 25: 1400.
  2. Moody GB, Muldrow WE, Mark RG (1984). "A noise stress test for arrhythmia detectors". Computers in Cardiology. 11: 381-384.
  3. Nemcova A, Smisek R, Opravilová K, Vitek M, Smital L, Maršánová L (2020). "Brno University of Technology ECG Quality Database (BUT QDB) (version 1.0.0)". PhysioNet.
  4. Moody G, Mark R (2001). "The impact of the MIT-BIH Arrhythmia Database". IEEE Engineering in Medicine and Biology Magazine. 20: 45–50.
  5. Goldberger AL, et al. (2000). "PhysioBank, PhysioToolkit, and PhysioNet: Components of a New Research Resource for Complex Physiologic Signals". Circulation. 101(23): e215–e220.
  6. Gilon C, Grégoire JM, Mathieu M, Carlier S, Bersini H (2023). "IRIDIA-AF, a large paroxysmal atrial fibrillation long-term electrocardiogram monitoring database". Scientific Data. 10: 714.
  7. Tsutsui K, et al. (2025). "SHDB-AF: a Japanese Holter ECG database of atrial fibrillation". Scientific Data. 12: 454.
  8. Petrutiu S, Sahakian AV, Swiryn S (2007). "Abrupt changes in fibrillatory wave characteristics at the termination of paroxysmal atrial fibrillation in humans". Europace. 9: 466-470.

Files

Total uncompressed size: 92.5 MB.

Access the files

Visualize waveforms

Folder Navigation: <base>
Name Size Modified
wfdb_files
ANNOTATORS (download) 83 B 2026-09-29
LICENSE.txt (download) 19.9 KB 2026-11-02
README.md (download) 3.3 KB 2026-09-29
RECORDS (download) 103.8 KB 2026-09-29
SHA256SUMS.txt (download) 2.1 MB 2026-11-02