I-CARE: International Cardiac Arrest REsearch consortium Database

Abstract

The International Cardiac Arrest REsearch consortium (I-CARE) Database includes baseline clinical information and continuous electroencephalography (EEG) recordings from comatose patients with cardiac arrest who were admitted to an intensive care unit in seven academic hospitals in the U.S. and Europe. Patients were monitored with 18 bipolar EEG channels from hours to several days for the diagnosis of seizures as well as for neurological prognostication. Long-term neurological function was determined using the Cerebral Performance Category scale.

Background

More than 6 million cardiac arrests happen every year worldwide, with survival rates ranging from 1% to 10% depending on geographic location [1]. Severe brain injury is the main determinant of poor outcome for patients surviving cardiac arrest resuscitation every year [1,2]. Most patients surviving to intensive care unit admission will be comatose, and 50-80% will have life-sustaining therapies withdrawn due to a perceived poor neurological prognosis [3]. 

Brain monitoring with electroencephalography aims to reduce the subjectivity in neurologic prognostication following cardiac arrest [4-9]. Clinical neurophysiologists have identified numerous patterns of brain activity that help to predict prognosis following cardiac arrest, including the presences of reduced voltage, burst suppression (alternating periods of high and low voltage), seizures, and a variety of seizure-like patterns [8]. The evolution of electroencephalogram (EEG) patterns over time may provide additional predictive information [6,7]. However, qualitative interpretation of continuous EEG is laborious, expensive, and requires review from neurologists with advanced training in neurophysiology who are unavailable in most medical centers.

Automated analysis of continuous EEG data has the potential to improve prognostic accuracy and to increase access to brain monitoring where experts are not readily available [6,7]. However, the datasets used in most studies typically only have small numbers of patients (<100) from single hospitals, which are unsuitable for deployment of several types of machine learning methods for EEG data analysis. To overcome this limitation the International Cardiac Arrest REsearch consortium (I-CARE) assembled a large collection of EEG data and neurologic outcomes from comatose patients who underwent EEG monitoring following cardiac arrest. The I-CARE dataset includes seven hospitals from the United States and Europe. 

Methods

The database originates from seven academic hospitals in the U.S. and Europe led by investigators part of the International Cardiac Arrest REsearch consortium (I-CARE).

1. Rijnstate Hospital, Arnhem, The Netherlands (Jeannette Hofmeijer).
2. Medisch Spectrum Twente, Enschede, The Netherlands (Barry J. Ruijter, Marleen C. Tjepkema-Cloostermans, Michel J. A. M. van Putten).
3. Erasme Hospital, Brussels, Belgium (Nicolas Gaspard).
4. Massachusetts General Hospital, Boston, Massachusetts, USA (Edilberto Amorim, Wei-Long Zheng, Mohammad Ghassemi, and M. Brandon Westover).
5. Brigham and Women’s Hospital, Boston, Massachusetts, USA (Jong Woo Lee).
6. Beth Israel Deaconess Medical Center, Boston, Massachusetts, USA (Susan T. Herman).
7. Yale New Haven Hospital, New Haven, Connecticut, USA (Adithya Sivaraju).

This database consists of clinical and EEG data from adult patients with out-of-hospital or in-hospital cardiac arrest who had return of heart function (i.e., return of spontaneous circulation [ROSC]) but remained comatose - defined as inability to follow verbal commands and a Glasgow Coma Score inferior or equal to 8. The initial database release contains data from 607 patients - this is the public training set for the 2023 PhysioNet Challenge. This database release does not contain data from the remaining 413 patients that we are retaining as the hidden validation and test sets for the Challenge.

All patients were admitted to an ICU and had their brain activity monitored with continuous EEG. Monitoring was typically started within hours of cardiac arrest and continues for several hours to days depending on the patients’ condition, so recording start time and duration varies from patient to patient. This database include EEG clips (up to 5-minute long) obtained hourly from the time of initiation of EEG and up to 72 hours from ROSC for each individual patient. This project includes a subset of the database that we have shared as a public training set for the George B. Moody PhysioNet Challenge 2023; the remaining part of the database has been retained as private validation and test sets for the Challenge.

Clinical Data

Patient information recorded at the time of admission (age, sex), location of arrest (out or in-hospital), type of cardiac rhythm recorded at the time of resuscitation (shockable rhythms include ventricular fibrillation or ventricular tachycardia and non-shockable rhythms include asystole and pulseless electrical activity), and the time between cardiac arrest and ROSC. Patient temperature after cardiac arrest is controlled using a closed-loop feedback device (TTM) in most patients unless there are contraindications such as severe and difficult to control hypotension or delay in hospital admission. For patients undergoing TTM, the temperature level can be controlled at either 36 or 33 degrees Celsius.

Neurological Prognostication and Outcome Assessment

All participating hospitals have protocols for multimodal neurological prognostication that follow international guideline recomendations. Formal neurological prognostication is deferred until the normothermia phase and confounding from sedatives can be minimized.

Patient Outcomes

Clinical outcome was determined prospectively in two centers by phone interview (at 6 months from ROSC), and at the remaining five hospitals retrospectively through chart review (at 3-6 months from ROSC). Neurological function was determined using the best Cerebral Performance Category (CPC) scale [10]. CPC is an ordinal scale ranging from 1 to 5, ranging from good neurological function to death.

De-identification

Clinical and EEG data were de-identified. Patients with age above 89 years old are listed with age "90". EEG timestamps are organized based on the time elapsed since ROSC.

Data Description

EEG Data

All EEG signal data is provided in WFDB format, with the signal data stored in MATLAB (MAT v4 format). MAT file. For example, the file named “ICARE_0284_06.mat” contains  EEG signal data from the sixth hour after cardiac arrest for patient "0284". The accompanying WFDB header file describes the contents of the WFDB signal file.

All EEG data were downsampled to 100 Hz. Each file contains an array with EEG signals from 18 bipolar channel pairs. Patients may have EEG started several hours after the arrest or need to have brain monitoring interrupted transiently while in the ICU, so gaps in data may be present. The EEG recordings continue for several hours to days, so the EEG signals are prone to quality deterioration from non-physiological artifacts. Each EEG file contains up to five minutes of EEG signal data. Most segments contain five minutes of signal data, however some might have less based on data availability for that hour. Only the cleanest 5 minutes of EEG data per hour are provided (this 5-minute epoch is selected based on artifact screenign described below). However, some EEG data might have one or many channels contaminated by artifact while some patients might only have artifactual, non-physiological data for the entire 5-minutes available.

Artfiact Screening (Signal Quality)

In addition to the EEG signal data, one additional file includes the artifact scores for each hour (e.g. “ICARE_0284.tsv”). This table contains the timestamp for the start of each EEG signal file in relation to the time of cardiac arrest (under the column “Time”). It also includes the quality of the EEG signal for that 5-minute epoch (column “Quality”). This artifact score is based on how many 10-second epochs within a 5-minute EEG window are contaminated by artifacts. Each 10-second epoch was scored for the presence of the following artifacts including: 1) flat signal, 2) extreme high or low values, 3) muscle artifact, 4) non-physiological spectra, and 5) implausibly fast rising or decreasing signal amplitude. Scores range from zero to one: a score of “1” indicates that all 5 minutes are free of artifact and a score of zero is entirely contaminated by artifact [6]. Automated artifact screening and generation of quality scores was performed without human validation, therefore user discretion is advised. Artifact scoring code is available as part of a pipeline hosted in Github under "Artifact_pipeline.zip" (https://github.com/bdsp-core/icare-dl) [11].

Clinical Data and Patient Outcome

The following clinical data is contained in each .txt file: 

Age (in years): number

Sex: Male, Female

ROSC (return of spontaneous circulation, in minutes): time from cardiac arrest to return of spontaneous circulation

OHCA (out-of-hospital cardiac arrest): True = out of hospital cardiac arrest: False = in-hospital cardiac arrest

VFib (ventricular fibrillation): True = shockable rhythm, False = non-shockable rhythm

TTM (targeted temperature management; in Celsius): 33, 36, or NaN for no TTM

Outcome: Good (CPC score of 1-2), Poor (CPC score of 3-5)

CPC: Cerebral Performance Category (CPC) score (ordinal scale 1-5)

CPC = 1: good neurological function and independent for activities of daily living 

CPC = 2: moderate neurological disability but independent for activities of daily living

CPC = 3: severe neurological disability

CPC = 4: unresponsive wakefulness syndrome [previously known as vegetative state] 

CPC = 5: dead. 

We have grouped CPC scores in two categories: 

• “Good outcome”: CPC = 1 or 2
• “Poor outcome”: CPC = 3, 4, or 5

Usage Notes

These data are in a WFDB-compatible format, and WFDB packages can be used to read them. 

We have implemented example prediction algorithms in MATLAB and Python that read the data:

MATLAB example at [12].

Python example at [13].

Release Notes

By downloading the data, you agree not to repost the data or to publish or otherwise share any work that uses the data, in full or in part, before the end of the PhysioNet Challenge 2023 except to the Computing in Cardiology conference.

Ethics

Data collection and analysis was performed under independent Institutional Review Board approvals at participating hospitals, and a data sharing agreement was made among participating hospitals. This was a retrospective analysis of data obtained as part of the usual care and the requirement for informed consent was waived.

Acknowledgements

This study was supported by the American Heart Association (20CDA35310297), CURE Epilepsy Foundation (Taking Flight Award), Neurocritical Care Society (NCS research training fellowship), Weil-Society of Critical Care Medicine Research Grant, the NIH (1K23NS090900, 1R01NS102190, 1R01NS102574, 1R01NS107291, 1K23NS119794), Epilepsiefonds (NEF 14-18), and Dutch Heart Foundation (2018T070).

Conflicts of Interest

M.V.P is founder of Clinical Science Systems. Clinical Science Systems did not contribute funding nor played any role in the study. M.B.W. is a co-founder of Beacon Biosignals. Beacon Biosignals did not contribute funding nor played any role in the study.

References

1. Yan, S., Gan, Y., Jiang, N. et al. The global survival rate among adult out-of-hospital cardiac arrest patients who received cardiopulmonary resuscitation: a systematic review and meta-analysis. Crit Care 24, 61 (2020). https://doi.org/10.1186/s13054-020-2773-2.
2. Dankiewicz J, Cronberg T, Lilja G, et al. Hypothermia versus Normothermia after Out-of-Hospital Cardiac Arrest. N Engl J Med. 2021;384:2283–2294.
3. Elmer J, Torres C, Aufderheide TP, Austin MA, Callaway CW, Golan E, Herren H, Jasti J, Kudenchuk PJ, Scales DC, Stub D, Richardson DK, Zive DM; Resuscitation Outcomes Consortium. Association of early withdrawal of life-sustaining therapy for perceived neurological prognosis with mortality after cardiac arrest. Resuscitation. 2016 May;102:127-35. doi: 10.1016/j.resuscitation.2016.01.016. Epub 2016 Feb 3. PMID: 26836944; PMCID: PMC4834233.
4. Amorim E, Rittenberger JC, Zheng JJ, et al. Continuous EEG monitoring enhances multimodal outcome prediction in hypoxic-ischemic brain injury. Resuscitation. 2016;109:121–126.
5. Hofmeijer J, Beernink TMJ, Bosch FH, Beishuizen A, Tjepkema-Cloostermans MC, van Putten MJAM. Early EEG contributes to multimodal outcome prediction of postanoxic coma. Neurology. 2015;85:137–143.
6. Zheng W-L, Amorim E, Jing J, et al. Predicting neurological outcome in comatose patients after cardiac arrest with multiscale deep neural networks. Resuscitation. 2021;169:86–94.
7. Zheng W-L, Amorim E, Jing J, et al. Predicting Neurological Outcome from Electroencephalogram Dynamics in Comatose Patients after Cardiac Arrest with Deep Learning. IEEE Trans Biomed Eng. Epub 2021.:1–1.
8. Khazanova D, Douglas VC, Amorim E. A matter of timing: EEG monitoring for neurological prognostication after cardiac arrest in the era of targeted temperature management. Minerva Anestesiol. 2021;87:704–713.
9. Ruijter BJ, van Putten MJAM, van den Bergh WM, Tromp SC, Hofmeijer J. Propofol does not affect the reliability of early EEG for outcome prediction of comatose patients after cardiac arrest. Clin Neurophysiol Off J Int Fed Clin Neurophysiol. 2019;130:1263–1270.
10. Taccone FS, Horn J, Storm C, et al. Death after awakening from post-anoxic coma: the “Best CPC” project. Crit Care Lond Engl. 2019;23:107.
11. International Cardiac Arrest EEG Consortium (ICARE) Dataset with Deep Learning. https://github.com/bdsp-core/icare-dl.
12. PhysioNet 2023 Challenge MATLAB Example. https://github.com/physionetchallenges/matlab-example-2023
13. PhysioNet 2023 Challenge Python Example. https://github.com/physionetchallenges/python-example-2023