PhysioNet/CinC Challenge 2014: Test Sets

Data used for the 2014 Challenge are 10-minute (or occasionally shorter) excerpts ("records") of longer multiparameter recordings of human adults, including patients with a wide range of problems as well as healthy volunteers. Each record contains four to eight signals; the first is an ECG signal in each case, but the others are a variety of simultaneously recorded physiologic signals that may be useful for robust beat detection. Signals have been digitized at rates between 120 and 1000 samples per second; in any given record, however, all signals are sampled at the same, fixed frequency.

A training data set for this challenge is available for study. It is a set of 100 records, named 100, 101, ..., 199, and it is provided in the set-p directory. You may wish to explore these records visually using LightWAVE. This data set is also available for download as a zip archive and as a tarball.

A new augmented training set, consisting of 100 records from the original \ test set is available as training.zip.The annotations were not generated from any specific channel and there was no fixed fiducial point, since some of the annotations were placed manually. The annotations include only beat labels and do not differentiate between beat types (all annotated beats were arbitrarily set to normal, 'N' beats).

The training set includes many records that can be processed without errors by the sample entry using the ECG only, but others will pose serious difficulty unless your entry makes good use of available information in the other signals; a few of the difficult records are 112, 133, 169, and 188.

A set of reference beat annotations for the training set is also available: set-p-atr.tar.gz. In this Challenge, reference beat annotations represent the preponderance of expert opinions about the locations of the observed (or imputed) QRS complexes in the ECG signal.

A separate hidden test data set was assembled for evaluating Challenge entries. Performance of the challenge entries on this hidden test set determined their rankings and thus the winners of the Challenge. The test set is not available here.

Important differences between the training set and the test set: The training set was intended to give participants an opportunity to see some of the problems their entries would face in the challenge, and to give us a way to verify that submitted entries are working as their authors intended. The performance of challenge entries on the training set did not contribute in any way to their scores and ranks in the Challenge.

The test set contains a wider variety of signals than in the training set. A successful entry needed to be able to discover their relationships and exploit features that can predict beat locations. Unlike the training set (sampled at a uniform 250 samples per second per signal), signals in the test set were sampled at rates between 120 and 1000 samples per second.

 Name                            Last modified      Size  Description
 Parent Directory                                     -   
 training.zip                    2015-01-16 14:59  113M  
 set-p2/                         2015-03-03 17:29    -   
 set-p/                          2014-01-07 19:18    -   
 set-p.zip                       2013-12-03 12:26  111M  
 set-p.tar.gz                    2013-12-03 12:25  111M  
 set-p-atr.tar.gz                2014-01-07 19:07   75K  
 DOI                             2015-09-21 13:00   19