VitalDB Arrhythmia Database: An Anesthesiologist-Validated Large-Scale Intraoperative Arrhythmia Dataset with Beat and Rhythm Labels

@article{PhysioNet-vitaldb-arrhythmia-1.0.0,
  author = {Eun, Dain and Shim, Kayoung and Lee, Hyunsoo and Lim, Yeji and Lim, Hanbyeol and Lee, Hyeonhoon and Lee, Hyung-Chul},
  title = {{VitalDB Arrhythmia Database: An Anesthesiologist-Validated Large-Scale Intraoperative Arrhythmia Dataset with Beat and Rhythm Labels}},
  journal = {{PhysioNet}},
  year = {2026},
  month = feb,
  note = {Version 1.0.0},
  doi = {10.13026/axd6-wm13},
  url = {https://doi.org/10.13026/axd6-wm13}
}

Abstract

Intraoperative cardiac arrhythmias present distinct characteristics and clinical challenges compared to non-surgical environments, yet publicly available electrocardiogram (ECG) databases have primarily focused on ambulatory and intensive care environments. To address this gap, we present the VitalDB Arrhythmia Database, a comprehensive collection of annotated intraoperative ECG recordings specifically designed for developing and validating arrhythmia detection algorithms in the surgical context. The database comprises 734,528 seconds of continuous ECG data from 482 surgical patients, with over 660,000 individually annotated heartbeats classified across four beat types and 10 distinct rhythm categories. To efficiently process the extensive source data, we developed a custom deep learning beat classifier that served as an automated screening tool for arrhythmia candidate segments. All annotations underwent rigorous validation by five anesthesiologists, with each segment independently reviewed by at least two anesthesiologists. Inter-rater reliability analysis demonstrated excellent agreement with an overall Cohen's kappa of 0.930 ± 0.130. This publicly accessible resource provides the research community with clinically validated intraoperative arrhythmia data, facilitating the development of robust detection algorithms suited to the unique physiological and technical challenges of the perioperative environment.

Background

The intraoperative period is characterized by a high incidence of cardiac arrhythmias triggered by surgical stimuli, anesthetic agents, and autonomic fluctuations [1, 2]. While intraoperative arrhythmias exhibit distinct patterns, existing public ECG datasets have predominantly consisted of data from ambulatory Holter recordings, failing to represent the unique characteristics of the surgical environment. The creation of such datasets has been challenging due to the need for experienced anesthesiologists to manually review hours of ECG recordings. While most perioperative arrhythmias are benign, they can require immediate intervention, and delayed treatment can lead to increased morbidity and mortality. The scarcity of labeled arrhythmia datasets from intraoperative ECG data has hindered the development of sophisticated detection algorithms. To address this, we analyzed 734,528 seconds of ECG data from VitalDB, a high-fidelity multi-parameter vital signs database in surgical patients [3, 4], with rhythm and beat labels validated by anesthesiologists, and present the VitalDB Arrhythmia Database as a publicly accessible dataset to facilitate future medical research and algorithm development in this unique clinical context.

Methods

The database was constructed through a multi-stage process involving automated screening of the public VitalDB open dataset [3, 4], followed by meticulous manual annotation and validation by five anesthesiologists.

1. Automated Candidate Screening: A deep learning beat classifier, UniMS-ECGNet, was used to efficiently identify potential arrhythmia segments. The algorithm was trained on data from the VitalDB open dataset, the MIT-BIH Arrhythmia Database [3, 5], and the CU Ventricular Tachyarrhythmia Database [3, 6]. Candidate segments were selected based on criteria such as consecutive abnormal beats, high R-R interval variability, or patterned ectopy. Further details on the model architecture and usage can be found in the referenced repository [7].

2. Annotation and Validation: Individual ECG beats were classified into four primary classes (Normal, Supraventricular, Ventricular, Unclassifiable). Each continuous segment was assigned one of ten final rhythm labels. All automated selections underwent rigorous clinical validation by a board of five anesthesiologists, with each segment being independently labeled by at least two physicians to reach a final consensus.

Data Description

The dataset includes a metadata file and 482 individual annotation files. The raw ECG waveform data is not included in this package but can be freely accessed and downloaded from the public VitalDB database using the corresponding case_id [3, 4]. Alternatively, the waveform data can be accessed programmatically using the VitalDB Python package (pip install vitaldb) without requiring manual file downloads or additional authentication.

• metadata.csv: A single file summarizing all cases, including columns like case_id, analyzed_duration_sec, total_beats, and a list of rhythm_classes present in the case. Additionally, it includes relevant surgical and clinical information.
• Annotation_Files/ folder: The folder contains a corresponding CSV file for each case, providing detailed beat and rhythm annotations. The annotation files are provided in the format Annotation_file_[case_id].csv for each case.

Column definitions for the annotation files are as follows:

• time_second: The timestamp of the R-peak in seconds, measured from the beginning of the recording.
• beat_type: The classification of the individual heartbeat.
• rhythm_label: The overall heart rhythm label for the segment in which the beat occurs.
• bad_signal_quality: A boolean marker (True/False) indicating if the beat is located within a segment of excessive noise or poor signal quality.
• bad_signal_quality_label: A label indicating the start or end of a bad signal quality segment (e.g., Start1, End1). This column is empty for rows not marking these specific boundaries.

Rhythm and Signal Quality Annotations

Rhythm Classes The database contains 10 distinct rhythm categories, with summary statistics presented below:

Signal Quality Labels Specific criteria were used to label segments with poor signal quality:

• Bad Signal Quality: This flag was used for segments where QRS complexes are visible but noise obscures P-wave or T-wave morphology, making accurate interpretation difficult. This label can be used alongside other rhythm labels when the underlying rhythm remains interpretable.
• Noise: This label was applied to segments where artifacts are so severe that QRS complexes themselves cannot be detected. Therefore, segments labeled as 'Noise' do not contain beat-level annotations.

Usage Notes

Accessing and Using the Data

The VitalDB Arrhythmia Database provides expert medical annotations for the ECG waveforms found in the comprehensive VitalDB, a high-fidelity multi-parameter vital signs database in surgical patients, which is also available on PhysioNet [3, 4].

While our annotations are specific to ECG arrhythmias, users can easily link them to the full, raw waveform data from the original VitalDB project. This allows researchers to seamlessly integrate our labels with synchronized biosignals like PPG and arterial pressure waveforms, paving the way for novel multi-modal algorithm development.

A detailed, executable guide on how to merge and utilize both datasets is provided in the UsageNote.ipynb Jupyter Notebook in our GitHub repository [8]. The recommended workflow is as follows:

1. Download Annotations

   Download the annotation files from this PhysioNet project.

2. Install Required Packages

   Install the necessary Python packages to access waveform data and handle annotations.

   pip install vitaldb pandas numpy matplotlib
   

3. Load ECG Waveforms and Annotations

   Use the case_id to download the ECG waveform from VitalDB and load the corresponding annotation file.

   import vitaldb
   import pandas as pd
   import numpy as np
   
   # Example case_id
   case_id = 337
   
   # Load ECG waveform from VitalDB (sampled at 500 Hz)
   vals = vitaldb.load_case(case_id, ['SNUADC/ECG_II'], 1/500)
   ecg_data = vals['SNUADC/ECG_II']
   
   # Load annotation file
   annotation_file = f'Annotation_file_{case_id}.csv'
   annotations = pd.read_csv(annotation_file)
   
   # Display the structure of annotations
   print(annotations.head())
   
   # Extract specific columns
   time_seconds = annotations['time_second'].values
   beat_types = annotations['beat_type'].values
   rhythm_labels = annotations['rhythm_label'].values
   signal_quality = annotations['bad_signal_quality'].values
   
   # Example: Get annotations for a specific time range (e.g., 100-110 seconds)
   start_time = 100
   end_time = 110
   segment_annotations = annotations[
       (annotations['time_second'] >= start_time) &
       (annotations['time_second'] <= end_time)
   ]
   
   print(f"\nAnnotations between {start_time}s and {end_time}s:")
   print(segment_annotations[['time_second', 'beat_type', 'rhythm_label']])
   
   # Example: Filter by specific rhythm type
   afib_beats = annotations[annotations['rhythm_label'] == 'Atrial fibrillation']
   print(f"\nTotal Atrial fibrillation beats: {len(afib_beats)}")
   

4. Combine and Analyze

   The annotation file contains beat-level information with timestamps (time_second), which can be matched with the ECG waveform data for visualization and analysis. Each time_second value corresponds to the R-peak location of a detected heartbeat.

Limitations

Release Notes

This is the initial release (version 1.0.0) of the VitalDB Arrhythmia Database.

Ethics

This project utilizes VitalDB, a high-fidelity multi-parameter vital signs database in surgical patients, a publicly available de-identified dataset from Seoul National University Hospital, approved by the institutional review board (H-1408-101-605). As this work involves secondary analysis of publicly accessible, de-identified data, no additional ethical approval was required.

All shared files are fully de-identified and contain no protected health information (PHI). To ensure privacy, no absolute dates or timestamps are included in the data, metadata, or filenames. All temporal information is provided as relative time offsets.

Conflicts of Interest

The authors declare no competing interests.

References

1. Kwon CH, Kim SH. Intraoperative management of critical arrhythmia. Korean J Anesthesiol. 2017;70(2):120-6.
2. Staikou C, Stamelos M, Stavroulakis E. Impact of anaesthetic drugs and adjuvants on ECG markers of torsadogenicity. Br J Anaesth. 2014;112(2):217-30.
3. Goldberger AL, Amaral LA, Glass L, Hausdorff JM, Ivanov PC, Mark RG, et al. PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals. Circulation. 2000;101(23):e215-20.
4. Lee HC, Park Y, Yoon SB, Yang SM, Park D, Jung CW. VitalDB, a high-fidelity multi-parameter vital signs database in surgical patients. Sci Data. 2022;9(1):279.
5. Moody GB, Mark RG. The impact of the MIT-BIH Arrhythmia Database. IEEE Eng Med Biol Mag. 2001;20(3):45-50.
6. Nolle FM, Badura FK, Catlett JM, Bowser RW, Sketch MH. CREI-GARD, a new concept in computerized arrhythmia monitoring systems. Comput Cardiol. 1986;13:515-8.
7. VitalDB. Vital beat noise detection [Internet]. GitHub; 2025 [cited 2026 Jan 24]. Available from: https://github.com/vitaldb/arrdb/blob/main/Vital_beat_noise_detection.ipynb
8. VitalDB. arrdb. GitHub [Internet]. [cited 2026 Jan 24]. Available from: https://github.com/vitaldb/arrdb/