PhysioNet/CinC Challenge 2015: Training Sets

This database holds the records used in the Physio\ Net/CinC Challenge 2015. See the page for more details.

Please cite the standard citation for PhysioNet when referencing this material:

Goldberger AL, Amaral LAN, Glass L, Hausdorff JM, Ivanov PCh, Mark RG, Mietus JE, Moody GB, Peng C-K, Stanley HE. PhysioBank, PhysioToolkit, and PhysioNet: Components of a New Research Resource for Complex Physiologic Signals. Circulation 101(23):e215-e220 [Circulation Electronic Pages; http://circ.ahajournals.org/cgi/content/full/101/23/e215]; 2000 (June 13).

Bedside monitor data leading up to a total of 1250 life-threatening arrhythmia alarms recorded from three of the largest intensive care monitor manufacturers’ bedside units will be used in this challenge. Data are sourced from four hospitals in the USA and Europe, chosen at random (and so do not necessarily represent the true statistics for false alarm rates for any given manufacturer or hospital which is likely to be different based on unit-specific protocols, software versions and unit types). The training and test sets have each been divided into two subsets of mutually exclusive patient populations. The training set contains 750 recordings and the test set contains 500 recordings. The test is unavailable to the public and will remain private for the purpose of scoring. No more than three alarms of each of the five categories are used from any given patient, and alarms are at least 5 minutes apart (usually longer). In this way, the competition does not address the issue of what to do with repeated alarms and how to use information from earlier alarms. Although this could be done with the full files from each patient, this may also a dangerous practice, because any algorithm would propagate and compound any errors from one alarm to the next. A team of expert annotators reviewed each alarm and labeled it either ‘true’, ‘false’, or ‘impossible to tell’. The Challenge includes only records that were reviewed by at least two annotators, of whom a two-thirds majority agreed that the alarm was either true or false.

An alarm was triggered 5 minutes from the beginning of each record. The exact time of the event that triggered the alarm varies somewhat from one record to another, but in order to meet the ANSI/AAMI EC13 Cardiac Monitor Standards [1], the onset of the event must be within 10 seconds of the alarm (i.e., between 4:50 and 5:00 of the record). Note that there may have been additional arrhythmia events in the 5 minutes preceding the alarm; these events have not been annotated.

In the “real-time” subset, each record is exactly five minutes long. In the “retrospective” subset (50% of the training set), each record contains an additional 30 seconds of data following the time of the alarm.

All signals have been resampled (using anti-alias filters) to 12 bit, 250 Hz and have had FIR band pass [0.05 to 40Hz] and mains notch filters applied to remove noise. Pacemaker and other nose artifacts may be present on the ECG. Pulsatile channels can suffer from movement artifact, sensor disconnects and other events (such as line flushes or coagulation in the catheter [2]). Each recording contains two ECG leads (which may or may not be the leads that triggered the alarm) and one or more pulsatile waveforms (the photoplethysmogram and/or arterial blood pressure waveform).

We have chosen not to provide the beat markers that the bedside alarm algorithms may have used to trigger the alarm. We also note that tachycardia and bradycardia alarms have variable thresholds, which are sometimes adjusted at the bedside. We have also removed these for consistency.

The 2015 Challenge data are provided in WFDB format with an Octave/MATLAB-compatible header. You can load the data directly using MATLAB or Octave without any special tools, but you must take care to scale it into the correct physical units. You can also use the WFDB MATLAB Toolbox function RDMAT ; to load data in physical units.

References

1. American National Standard (ANSI/AAMI EC13:2002) Cardiac monitors, heart rate meters, and alarms. Arlington, VA: Association for the Advancement of Medical Instrumentation; 2002.

2. Li Q, Mark RG, Clifford GD. Artificial arterial blood pressure artifact models and an evaluation of a robust blood pressure and heart rate estimator. Biomed Eng Online 2009;8:13

Icon  Name                            Last modified      Size  Description
[PARENTDIR] Parent Directory - [   ] training.zip 2015-03-06 17:29 323M [DIR] training/ 2015-03-06 17:12 -

Questions and Comments

If you would like help understanding, using, or downloading content, please see our Frequently Asked Questions.

If you have any comments, feedback, or particular questions regarding this page, please send them to the webmaster.

Comments and issues can also be raised on PhysioNet's GitHub page.

Updated Friday, 28 October 2016 at 16:58 EDT

PhysioNet is supported by the National Institute of General Medical Sciences (NIGMS) and the National Institute of Biomedical Imaging and Bioengineering (NIBIB) under NIH grant number 2R01GM104987-09.