Kingston ICU AF Dataset

@article{PhysioNet-kingston-icu-af-1.0.0,
  author = {Nassar, Sarah and Maghsoodi, Nooshin and Chen, Ziqi and Hamilton, Alexander and Maslove, David and Sibley, Stephanie and Laird, Phil and Mousavi, Parvin},
  title = {{Kingston ICU AF Dataset}},
  journal = {{PhysioNet}},
  year = {2026},
  month = jul,
  note = {Version 1.0.0},
  doi = {10.13026/47wc-qk15},
  url = {https://doi.org/10.13026/47wc-qk15}
}

Abstract

This newly published dataset aims to facilitate research in atrial fibrillation (AF) detection in the intensive care unit (ICU). The ICU is a unique environment with noisy signals due to patient movement and prevalent alarms that can lead to alarm fatigue in care providers. The dataset contains 596 labelled 10-second electrocardiograms (ECGs), with 99 atrial fibrillation/flutter samples and 497 sinus rhythm (regardless of the underlying heart rate) samples. These ECGs were collected from bedside monitors in the ICU of the Kingston General Hospital (KGH) in Kingston, Ontario, Canada. Each ECG contains four leads (I, II, III and a V1 equivalent) and was labelled by at least two clinicians.

Background

Atrial fibrillation (AF) is the most common cardiac arrhythmia and can lead to negative health outcomes such as heart failure and stroke [1]. The prevalence of AF is higher in the intensive care unit (ICU) than in the general population at up to 15-20% or more [2-4]. AF is primarily diagnosed by visually inspecting the electrocardiogram (ECG) reading of a patient and identifying morphological irregularities. AF management in ICU patients presents a unique challenge as these patients are at higher risk of rapid health deterioration. However, ICU patients are connected to bedside monitors that continuously capture their ECG readings, allowing for automatic monitoring of their cardiac rhythm.

Accurate detection of AF is crucial for patient health outcomes. A limitation in the current literature is the absence of comprehensive comparisons of approaches such as classical machine learning, deep learning, and foundation models, especially for data from ICU patients. This dataset, along with our performance benchmarks, will enable the research community to continue advancing the state-of-the-art in this field.

The MIMIC-IV-ECG database [5], which contains about 800,000 ECGs from nearly 160,000 ICU patients with more than 600,000 cardiologist reports, has yet to publish the de-identified free-text cardiologist reports. Therefore, to our knowledge, there is no other open-source dataset with expert-labelled ECGs from ICU patients.

Methods

Our institutional ICU dataset was collected from the Kingston General Hospital (KGH) in Kingston, Ontario, Canada. It includes archival data between 2015 and 2020 from bedside monitors covering 1,043 patients. ECG signals include four leads (I, II, III and a V1 equivalent) sampled at a frequency of 240 Hz.

From a subset of these patients, a randomly selected 10-second ECG recording per patient was labelled by two critical care physicians. A third physician was consulted in the case of ties. ECG recordings with three or fewer R peaks detected in any lead with the SleepECG Python package were dropped.

Out of the 596 labelled ECGs from 596 unique patients, 497 were labelled as sinus rhythm (regardless of the underlying heart rate), encoded as SINUS, and 99 were labelled as atrial fibrillation (combined with atrial flutter), encoded as AFIB/AFLT.

More details on the data collection and annotation process are described by Chen [6].

Data Description

Raw ECG waveforms can be found in WFDB format under the ECGs/ folder. There are a total of 596 ECGs, with each having a recoded ID between 0 and 595 (inclusive). ECGs are 10 seconds long with a sampling frequency of 240 Hz. The labels (SINUS and AFIB/AFLT) are provided in metadata.csv. There are 497 SINUS ECGs and 99 AFIB/AFLT ECGs. The breakdown of files is as follows:

• ECGs/ is the folder containing ECG waveform files. Each record has a numeric ID ranging from 0 to 595. ECGs contain four leads (I, II, III, and a V1 equivalent) sampled at 240 Hz.
• RECORDS is the file containing a list of paths to all ECG records.
• metadata.csv is the metadata file and contains three columns:
  1. ECG for each record path from RECORDS.
  2. Rhythm for binary rhythm class annotations (SINUS for sinus rhythm regardless of the underlying heart rate and AFIB/AFLT for atrial fibrillation or atrial flutter).
  3. TrainOrTest for the train/test split we used in our comparative study. This same split can be used for direct performance comparisons and benchmarking.

Usage Notes

This data can be used for atrial fibrillation detection research in the unique ICU setting. A comprehensive comparison of data-driven AI approaches, encompassing feature-based classifiers, deep learning, and foundation models, with reported benchmarks on this dataset has been published by Nassar et al. [7]. Future work can explore methods to improve detection performance despite the relatively small sample size, class imbalance, and noise arising from the ICU environment, which may affect generalizability.

ECG records can be loaded with the WFDB module in Python:

import wfdb
ecg = wfdb.rdrecord('ECGs/ECG_0').to_dataframe().values

The CSV metadata file containing labels can be loaded with the pandas module in Python:

import pandas as pd
metadata = pd.read_csv('metadata.csv')
record_paths = metadata['ECG']
rhythms = metadata['Rhythm']
train_or_test = metadata['TrainOrTest']

Release Notes

Version 1.0.0: Initial public release.

Ethics

The Queen’s University Health Sciences and Affiliated Teaching Hospitals Research Ethics Board provided ethics approval for the collection and publication of this data (protocol number: DMED-1863-15). The need for informed consent was waived because the data was already being collected as part of routine clinical practice and stored in de-identified format.

Conflicts of Interest

This work was supported in part by the Natural Sciences and Engineering Research Council of Canada (NSERC), Social Sciences and Humanities Research Council (SSHRC), and Vector Institute. S. Sibley, D. Pichora, and D. Maslove are members of the Attending Staff at KHSC. S. Sibley is supported in part by the Canadian Institutes of Health Research (CIHR), has received honoraria from Think Research, meeting sponsorships from Boston Scientific, Trimedic, and Icentia, and serves as a hospital organ donation physician with the Trillium Gift of Life Network at Ontario Health. D. Maslove is supported in part by the Southeastern Ontario Academic Medical Association (SEAMO). P. Mousavi is supported in part by a Canada CIFAR AI Chair and a Canada Research Chair.

References

1. Kornej J, Börschel CS, Benjamin EJ, Schnabel RB. Epidemiology of atrial fibrillation in the 21st century: novel methods and new insights. Circulation research. 2020 Jun 19;127(1):4-20.
2. Moss TJ, Calland JF, Enfield KB, Gomez-Manjarres DC, Ruminski C, DiMarco JP, Lake DE, Moorman JR. New-onset atrial fibrillation in the critically ill. Critical care medicine. 2017 May 1;45(5):790-7.
3. Paula SB, Oliveira A, e Silva JM, Simões AF, Gonçalves-Pereira J, Oliveira AC, e Silva JP, Simões Sr AF, Pereira JG. Atrial fibrillation in critically ill patients: Incidence and outcomes. Cureus. 2024 Feb 28;16(2).
4. Rottmann FA, Abraham H, Welte T, Westermann L, Bemtgen X, Gauchel N, Supady A, Wengenmayer T, Staudacher DL. Atrial fibrillation and survival on a medical intensive care unit. International Journal of Cardiology. 2024 Mar 15;399:131673.
5. Gow B, Pollard T, Nathanson LA, Johnson A, Moody B, Fernandes C, et al. MIMIC-IV-ECG: Diagnostic Electrocardiogram Matched Subset. Version 1.0. PhysioNet; 2023. doi:10.13026/4nqg-sb35.
6. Chen B. Deep Learning-Driven Identification of Atrial Fibrillation in the ICU. Queen's University (Canada); 2021.
7. Nassar S, Maghsoodi N, Mannina S, Addas S, Sibley S, Fichtinger G, Pichora D, Maslove D, Abolmaesumi P, Mousavi P. A Dataset and Benchmarks for Atrial Fibrillation Detection from Electrocardiograms of Intensive Care Unit Patients. IEEE Trans Biomed Eng. 2026 Jul 20. doi:10.1109/TBME.2026.3715145.