SHDB-AF: a Japanese Holter ECG database of atrial fibrillation

Abstract

Saitama Heart Database Atrial Fibrillation (SHDB-AF) is a novel open-sourced Holter ECG database from Japan, containing data from 100 unique patients with paroxysmal atrial fibrillation. The dataset contains raw ECG recordings with manually annotated rhythm at the beat level performed by a fellow in cardiology. The dataset was collected as part of a study evaluating the generalization performance of a deep learning atrial fibrillation event detection model across different distribution shifts.

Background

Atrial fibrillation (AF) is a prevalent atrial arrhythmia that reduces quality of life and leads to complications such as embolic stroke and heart failure. Recent progress in machine learning and deep learning (DL) has demonstrated the potential to improve diagnostic accuracy significantly. Ensuring that DL models are robust and applicable across diverse factors such as ethnicity, age, and sex is crucial.

Despite the availability of several ECG databases to the research community, such as MITDB [1, 2], AFDB [2, 3], LTAFDB [2, 4], IRIDIA-AF [5], Icentia11k [2, 6, 7], and CPSC2021 [2, 8], none of them incorporate a sample from the Japanese population.

The SHDB-AF dataset was collected in previous work to evaluate the generalization performance of a DL algorithm for AF event detection from beat-to-beat time intervals, termed ArNet2 [9, 10], under different distribution shifts. A second DL model named RawECGNet [11] was developed and benchmarked against ArNet2. In contrast to ArNet2, RawECGNet was developed based on the raw, single-lead ECG signal. Regarding F1 score, ArNet2 achieved an F1 score of 0.92 for SHDB using beat-to-beat time intervals and 0.93 for RawECGNet using the raw ECG signal. Further results are presented in associated papers [10, 11].

Methods

As previously mentioned, data provided in this project was initially collected for the purpose of evaluating generalization performance of a DL algorithm for AF detection denoted ArNet2, [9, 10]. Inclusion criteria and preprocessing steps applied to SHDB-AF, are described bellow.

Data collection

This database includes ECG recordings of adult patients who underwent Holter monitoring as ordered by their treating physician from November 2019 to January 2022. Holters were recorded using Fukuda Holter monitor and digitized at 125Hz with two leads recorded, modified CC5 and NASA leads. Each recording lasts approximately 24 hours. While no reference beat annotations were extracted, each recording includes a diagnosis based on the free-text medical report prepared following the patient's examination.

Data preparation

Inclusion criteria

In total, 147 Holter recordings were collected, from which 100 recordings belonging to 100 unique patients were selected. Following the approach used in Biton et al. [10], 100 recordings from each respective database underwent re-annotation by a fellow in cardiology (MA). These subsets were stratified according to age, sex, and AF diagnosis. Specifically, 80 recordings were from patients diagnosed with AF based on the per-recording diagnosis. Further details regarding the inclusion criteria for these selected recordings are outlined in Biton et al. [10].

preprocessing

Subsequently, all recordings underwent filtering using a zero-phase second-order infinite impulse response bandpass filter with a passband of [0.67 - 100] Hz [12] to eliminate baseline wander and high-frequency noise. Following this, the recordings were resampled to 200 Hz using an anti-aliasing filter. Beat annotations were then identified using the epltd implementation of the Pan and Tompkins algorithm [13].

Annotation protocol

All recordings underwent manual beat-level annotation. To ensure patient confidentiality under HIPAA/GDPR regulations, cardiologists accessed a secure server remotely for the annotation process. Specifically, the PhysioZoo software [14, 15] was used to annotate the recordings. An annotation protocol was established, focusing on categorizing supraventricular arrhythmias including (1) AF, (2) Atrial Flutter, (3) Atrial tachycardia, and (4) other supraventricular tachycardias such as Wolf-Parkinson-White and intranodal tachycardias. Normal sinus rhythm and other rhythms were not annotated. MA spent an estimated average of 45 minutes per 24-hour Holter recording for annotation. A detailed description of the re-annotation protocol can be found in Biton et al. [10].

De-identification

Patient identifiers were anonymized following HIPAA guidelines and local regulations. Each participant was assigned a random 3-digit unique identifier ranging from 000 to 100. Numbers with fewer than three digits were padded with zeros to maintain consistent length. Dates directly associated with participants were omitted, retaining only the year of recording and the relative time of recording in the day.

Data Description

ECG data

All ECG recordings are stored in the WFDB format. Each ECG file (.dat suffix) contains two channels: 'ECG1', representing modified CC5 leads, and 'ECG2', representing NASA leads. The 'base_year' field denotes the year of recording, and 'base_time' indicates the relative start time within the day.

Rhythm annotations

Rhythm annotations are provided at the beat level in separate annotation files (.atr). These files categorize rhythms into five types:

1. (AFIB (atrial fibrillation);
2. (AFL (atrial flutter);
3. (AT (atrial tachycardia);
4. (PAT (Other supraventricular tachycardias such as Wolf-Parkinson-White) and (NOD (intranodal tachycardias); and
5. (N (other, such as NSR, that were not labeled).

In the annotation files, rhythm annotations are listed under the 'aux_note' parameter. Each rhythm mark within the annotation files indicates the start of a rhythm interval. The appearance of a subsequent rhythm mark marks the beginning of a new interval.

The shortest interval lasts 2.5 seconds, while the longest extends to 24 hours, with a median duration of 47.5 seconds (Q2-Q3 range: 17.0 to 270.25 seconds).

Below are the counts for each annotated beat by rhythm and the overall intervals.

R-peak annotations are provided in files with the .qrs suffix.

Clinical data

Furthermore, the AdditionalData.csv file provides detailed clinical and demographic information, including those key attributes for each record:

• <Study ID>: A unique three-digit identifier.
• <Age>: The subject's age in years at the time of the record.
• <Sex>: The subject's gender, categorized as "male" or "female".
• <Dx>: The final diagnosis based on the medical report prepared following the Holter examinations.
• <AFL>: An indicator of whether atrial flutter (AFL) was present in the Holter recording.
• <Previous ablation>: An indicator of whether the patient had undergone catheter ablation prior to the Holter examination.
• <Pacemaker>: An indicator of whether the subject had a pacemaker implanted during the Holter recording.
• <AAD>: Any permanent medication taken by the subject prior to the Holter recording.

Each file name corresponds uniquely to a study ID, specifically matching the 'Study ID' field in the AdditionalData.csv.

The file included, example.png, provides an illustration of an ECG example.

Usage Notes

This dataset has been used in the publications "Generalizable and robust deep learning algorithm for atrial fibrillation diagnosis across geography, ages, and sexes" and "RawECGNet: Deep Learning Generalization for Atrial Fibrillation Detection From the Raw ECG". It can be used as a benchmark dataset suitable for developing and validating arrhythmia detection algorithms. Further details about the dataset can be found in our associated paper [16].

Release Notes

Version 1.0.0: initial release.

Ethics

The authors declare no ethics concerns. The present work was approved by the institutional ethics committee at Saitama Medical University International Medical Center (IRB number 2023-145). Due to retrospective and descriptive nature of the study, written informed consent was waived.

Acknowledgements

Hittman: Technion EVPR Fund: Hittman Family Fund.

Conflicts of Interest

The authors have no conflicts of interest to declare.

References

1. Moody GB, Mark RG. The impact of the MIT-BIH arrhythmia database. IEEE engineering in medicine and biology magazine. 2001 May;20(3):45-50.
2. Goldberger AL, Amaral LA, Glass L, Hausdorff JM, Ivanov PC, Mark RG, Mietus JE, Moody GB, Peng CK, Stanley HE. PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals. circulation. 2000 Jun 13;101(23):e215-20.
3. Moody G. A new method for detecting atrial fibrillation using RR intervals. Proc. Comput. Cardiol.. 1983;10:227-30.
4. Petrutiu S, Sahakian AV, Swiryn S. Abrupt changes in fibrillatory wave characteristics at the termination of paroxysmal atrial fibrillation in humans. Europace. 2007 Jul 1;9(7):466-70.
5. Gilon C, Grégoire JM, Mathieu M, Carlier S, Bersini H. IRIDIA-AF, a large paroxysmal atrial fibrillation long-term electrocardiogram monitoring database. Scientific data. 2023 Oct 18;10(1):714.
6. Tan, S., Ortiz-Gagné, S., Beaudoin-Gagnon, N., Fecteau, P., Courville, A., Bengio, Y., & Cohen, J. P. (2022). Icentia11k Single Lead Continuous Raw Electrocardiogram Dataset (version 1.0). PhysioNet.
7. Tan S, Androz G, Chamseddine A, Fecteau P, Courville A, Bengio Y, Cohen JP. Icentia11k: An unsupervised representation learning dataset for arrhythmia subtype discovery. arXiv preprint arXiv:1910.09570. 2019 Oct 21.
8. Wang X, Ma C, Zhang X, Gao H, Clifford GD, Liu C. Paroxysmal atrial fibrillation events detection from dynamic ecg recordings: The 4th china physiological signal challenge 2021. Proc. PhysioNet. 2021:1-83.
9. Chocron A, Oster J, Biton S, Mandel F, Elbaz M, Zeevi YY, Behar JA. Remote atrial fibrillation burden estimation using deep recurrent neural network. IEEE Transactions on Biomedical Engineering. 2020 Dec 4;68(8):2447-55.
10. Biton S, Aldhafeeri M, Marcusohn E, Tsutsui K, Szwagier T, Elias A, Oster J, Sellal JM, Suleiman M, Behar JA. Generalizable and robust deep learning algorithm for atrial fibrillation diagnosis across geography, ages and sexes. NPJ Digital Medicine. 2023 Mar 17;6(1):44.
11. Ben-Moshe N, Tsutsui K, Biton S, Zvuloni E, Sörnmo L, Behar JA. RawECGNet: Deep Learning Generalization for Atrial Fibrillation Detection From the Raw ECG. IEEE Journal of Biomedical and Health Informatics. 2024 May 24.
12. Kligfield P, Gettes LS, Bailey JJ, Childers R, Deal BJ, Hancock EW, Van Herpen G, Kors JA, Macfarlane P, Mirvis DM, Pahlm O. Recommendations for the standardization and interpretation of the electrocardiogram: part I: the electrocardiogram and its technology: a scientific statement from the American Heart Association Electrocardiography and Arrhythmias Committee, Council on Clinical Cardiology; the American College of Cardiology Foundation; and the Heart Rhythm Society endorsed by the International Society for Computerized Electrocardiology. Circulation. 2007 Mar 13;115(10):1306-24.
13. Pan J, Tompkins WJ. A real-time QRS detection algorithm. IEEE transactions on biomedical engineering. 1985 Mar(3):230-6.
14. Behar JA, Levy J, Zvuloni E, Gendelman S, Rosenberg A, Biton S, Derman R, Sobel JA, Alexandrovich A, Charlton PH, Goda MÁ. PhysioZoo: The Open Digital Physiological Biomarkers Resource. In2023 Computing in Cardiology (CinC) 2023 Oct 1 (Vol. 50, pp. 1-4). IEEE.
15. Gendelman S, Biton S, Derman R, Zvuloni E, Levy J, Lugassy S, Alexandrovich A, Behar JA. PhysioZoo ECG: Digital electrocardiography biomarkers to assess cardiac conduction. In2021 Computing in Cardiology (CinC) 2021 Sep 13 (Vol. 48, pp. 1-4). IEEE.
16. Tsutsui K, Brimer SB, Ben-Moshe N, Sellal JM, Oster J, Mori H, Ikeda Y, Arai T, Nakano S, Kato R, Behar JA. SHDB-AF: a Japanese Holter ECG database of atrial fibrillation. arXiv preprint arXiv:2406.16974. 2024 Jun 22.