20 March: New questions and answers in the FAQ.
2 April: Test set B (100 records) has been posted, and 50 additional records with annotations have been added to training set A. Details and links are below.
30 April: Event 6 has been cancelled (see below).
2 August: Phase 2 is now open. See How to submit a Challenge 2013 entry for information about the required contents and acceptable formats for entries. Each member of the PhysioNet/CinC Challenge 2013 project now has a personal participant page; go there for additional information and to submit your entries.
22 September: Slides shown by the participants at the CinC sessions devoted to the Challenge are now available.
24 October: Top scores have been posted, following the conclusion of the Challenge.
9 December: Participants' CinC 2013 papers about the Challenge and the top open-source entries have been posted.
If you missed the Challenge deadlines, you are still welcome to participate unofficially (without eligibility for awards). To begin, make a PhysioNetWorks account for yourself if you don't already have one, then join the PhysioNet/CinC Challenge 2013 project on PhysioNetWorks.
Since the late 19th century, decelerations of fetal heart rate have been known to be associated with fetal distress. Intermittent observations of fetal heart sounds (auscultation) became standard clinical practice by the mid-20th century. The first fetal heart rate (FHR) monitors were developed more than 50 years ago, and became widely available by the mid-1970s. Continuous FHR monitoring was expected to result in dramatic reduction of undiagnosed fetal hypoxia, but disillusionment rapidly set in as studies showed that the outputs of FHR monitors were often unreliable and difficult to interpret, resulting in increased rates of Caesarean deliveries of healthy infants, with little evidence that reductions in adverse outcomes were attributable to the use of FHR monitors (see FHR History).
Improved accuracy in FHR estimation has been achieved through use of more sophisticated signal processing techniques applied to more reliable signals. These improvements, coupled with a better understanding of the limitations of fetal monitoring, have led to wider acceptance. There remains a great deal of room for improvement, however. The most accurate method for measuring FHR is direct fetal electrocardiographic (FECG) monitoring using a fetal scalp electrode. This is possible only in labor, however, and is not common in current clinical practice because of its associated risks.
Noninvasive FECG monitoring makes use of electrodes placed on the mother's abdomen. This method can be used throughout the second half of pregnancy and is of negligible risk, but it is often difficult to detect the fetal QRS complexes in ECG signals obtained in this way, since the maternal ECG is usually of greater amplitude in them.
Figure 1. Four simultaneous noninvasive fetal ECG signals, acquired using electrodes placed on the mother's abdomen, containing both the fetal and the maternal ECGs.. The maternal QRS complexes (not marked) are larger than the fetal QRS complexes (marked in blue). The figure is interactive; click on the large blue circle to remove and restore the marker bars. Click on the Help tab for information about other options.
Other features of the direct fetal ECG, such as FHR variability and fetal QT interval, may be useful independent indicators of fetal status. There are no accepted techniques for assessing such features from noninvasive FECG, however. Before such techniques can be developed, it will be necessary to establish accurate methods for locating QRS complexes and for estimating the QT interval in noninvasive FECG.
The aim of this year's PhysioNet/Computing in Cardiology Challenge is to encourage development of accurate algorithms for locating QRS complexes and estimating the QT interval in noninvasive FECG signals. Using carefully reviewed reference QRS annotations and QT intervals as a gold standard, based on simultaneous direct FECG when possible, the challenge is designed to measure and compare the performance of participants' algorithms objectively. Multiple challenge events are designed to test basic FHR estimation accuracy, as well as accuracy in measurement of inter-beat (RR) and QT intervals needed as a basis for derivation of other FECG features.
Data for the challenge consist of a collection of one-minute fetal ECG recordings. Each recording includes four noninvasive abdominal signals as illustrated above (figure 1). The data were obtained from multiple sources using a variety of instrumentation with differing frequency response, resolution, and configuration, although in all cases they are presented as 1000 samples per signal per second.
In each case, reference annotations marking the locations of each fetal QRS complex were produced, usually with reference to a direct FECG signal, acquired from a fetal scalp electrode. The direct signals are not included in the challenge data sets, however.
As in recent challenges, the data have been divided into three sets:
Note that the training data set (set A) does not include examples of all of the sources included in the test sets. It will be necessary to design methods for addressing the challenge that are sufficiently flexible to work with data that have recording characteristics that are similar, but not identical, to those in the training set. This feature of the challenge data is intended to encourage participants to develop approaches that are compatible with the variety of devices and signals encountered in clinical practice, and to allow us to assess how successfully participants have been able to accomplish this goal.
Reference QT interval measurements are in preparation; those for records in set A will be made available to participants when they are complete.
Using the direct FECG when possible, reference annotations marking the locations of the fetal QRS complexes (denoted by N) have been derived by crowd-sourcing using a mixture of experts, volunteers, and algorithms.
These reference annotations have been processed to derive the reference RR interval time series (the intervals between successive N annotations).
The reference RR interval time series have been processed further to derive the reference heart rate (HR) time series. Each HR measurement is determined by a widely-used IPFM-based method over a 6-second window (i.e., by the number of reference RR intervals falling within a 6-second interval, including fractional intervals that fall only partly within the interval). Successive windows overlap by 50% (3 seconds), but the first and last windows in each one-minute record are omitted (since the lengths of the initial and final RR interval in the record are unknown), and any windows that include a gap in the RR interval series are likewise omitted. Thus there are at most 16 HR measurements per one-minute record.
A single reference QT interval measurement is produced for each record for which the direct ECG is available. Using the reference QRS annotations to time-align the direct fetal ECG waveforms, the median fetal cardiac cycle is computed, and the reference QT interval is measured manually by inspection of the median cycle. The software used to compute the median cycle is available in source form to participants.
Participants create software that annotates a challenge record with reference to the noninvasive ECG signals only. Important: Software created for the challenge should annotate only one record and then exit. In the challenge, it should not be assumed that any other records are available for reference at run-time.
An entry must produce a test annotation file for any record presented to it, in the same format as the reference annotation files. (Either the PhysioBank-compatible binary format, as in the *.fqrs files provided for set A, or the text format, as in the *.fqrs.txt files for set A, is acceptable.) The challenge organizers will derive test RR and HR time series from the test annotation files, using the same software that was used to derive the reference RR and HR time series. This software is available for participants' use in development and self-evaluation of their challenge entries.
Optionally, participants' algorithms may compute a test QT interval measurement for each record, and append that measurement to the test annotation file as shown in the example. It may be advantageous to make such a measurement using a method similar to that used for the reference QT interval measurements, but participants are free to use any method of their choice. Note that QT interval durations will vary among the four noninvasive signals; the challenge is to match (as closely as possible) the reference QT interval measured from the direct signal, which is not available except in the learning set. This will not be easy!
This year's challenge is structured as three pairs of events. Events 1, 2, and 3 are the major events for which monetary awards will be given to the top-ranked participants; they require submission of open-source entries to be tested by the challenge organizers using test set C. Events 4, 5, and 6 are minor events in which the participants submit their entries' annotation files for each record in test set B; non-monetary awards will be given to the top-ranked participants in these events.
Scores in these events are computed from the differences between matching reference and test FHR measurements, calculated over the entire test set (C for event 1, B for event 4).
Scores in these events are computed from the differences between matching reference and test RR intervals, calculated over the entire test set (C for event 2, B for event 5).
Scores in this event are calculated from the differences between matching reference and test QT intervals, calculated over the test set C. Since direct ECG signals, hence reference QT intervals, are not available for all records in the test set, some records will not be scored. Entries should make test QT interval measurements for all records if their authors wish to participate in event 3, since we will not publish a list of records that will be scored for this event.
We originally planned and announced an event 6 for QT interval measurement on test set B. This event has been cancelled. Reliable reference fetal QT measurements are available for only a subset of challenge records, and we planned to score events 3 and 6 using only these records. Since there were fewer such records than anticipated, it was necessary to allocate these records to sets A and C only; hence we are unable to score event 6 since there are no reliable means of determining reference QT intervals for the set B records.
To begin, we recommend studying the learning set as preparation for the Challenge itself. A sample entry that can be used as a model for your entries, and software for scoring your entries using the learning set, are available here (in preliminary versions, physionet2013.m and genresults.m, for MATLAB). Scores obtained from learning set data are not used for ranking entries, however! We provide the scoring software so that participants can verify that they are able to prepare properly formatted entries. In general, participants should expect that official scores obtained using the test sets will differ from unofficial scores obtained using the learning set, especially if the algorithms have been (over)trained.
All entries must include (or be able to produce) annotation files in the format of the reference annotation files provided for set A.
Open-source entries must include the sources for the software used to produce the annotations. As in the sample entries, your entry must write the annotations to standard output rather than to a file; our test framework captures the standard output of your entry for processing. Your entry may be written in portable (ANSI/ISO) C or MATLAB/Octave m-code; other languages, such as C++, Java, Perl, Python, and R, may be acceptable, and your entry may make use of open-source libraries that are available for Linux, but please ask us first, and do so no later than 7 April 2013.
Participants submitting open-source entries are entered into events 1 and 2, and into event 3 if their entries include QT measurements. These entries will be run on set C by the challenge organizers; scores will be returned as soon as possible after submission, but participants should expect that scoring will require a day or two (possibly longer near the major deadlines).
Participants submitting set B annotations ("closed-source entries") are entered into events 4 and 5, and into event 6 if their entries include QT measurements.
Participants may submit both open-source entries and closed-source entries, but the total number of entries from any participant or team is limited to 8 (3 in Phase 1, and 5 in Phase 2, as described below).
Awards will be presented to the most successful eligible participants during Computing in Cardiology (CinC) 2013. To be eligible for an award, you must:
An important goal of this Challenge, and of others in the annual series of PhysioNet/CinC Challenges, is to accelerate progress on the Challenge questions, not only during the limited period of the Challenge, but also afterward. In pursuit of this goal, we strongly encourage participants to submit open-source entries that will be made freely available after the conclusion of the Challenge via PhysioNet.
Eligible authors of the entries that receive the best test set scores in each Challenge event will receive award certificates during the closing plenary session of CinC on 25 September 2013. In recognition of their contributions to further work on the Challenge problems, eligible authors of the open-source entries that receive the best test set scores will also receive monetary awards. No team or individual will receive more than one such monetary award.
No. We are trying to encourage both experimentation with multiple approaches and sustained effort. In past Challenges some participants have used their entire allowance of entries before the first deadline, and others have saved their entries until hours before the final deadline. The most successful participants have usually reflected on each set of results, refining their ideas (and not merely their decision thresholds) before submitting the next entry. This approach yields better results, and it also allows us to review your entries and provide scores more rapidly than if we receive a large fraction of them just before the deadlines.
The output for each 1-minute test record should be a file of annotations.
If you enter the open-source events (1, 2, and 3), your entry must be in the form of software. We will run your software on each of the records in the hidden test set (set C), and redirect its standard output to a file for each test record. The requirement to write to standard output rather than to a named file allows us to run your software in a secure "sandbox" environment, which is necessary to make it practical for us to conduct these events.
If you enter the other events (4, 5, and 6), your entry should be a set of annotation files that you will create yourself by running your software using each of the records in the open test set (set B). Since you will run your software yourself for these events, it may create the files directly, or (as for the open-source events) it can write the annotations to its standard output and you can capture that standard output in a file.
See What is a "standard input" or a "standard output"? if this is confusing.
An additional point of potential confusion is that we have provided reference annotation files for the training set (set A) in both .fqrs (PhysioBank- compatible binary) and .fqrs.txt (text) formats. We expect that most participants will find it easier to generate text-format output, but we will accept either format.
There are no MS Excel files in the challenge data sets.
The .fqrs.txt files contain times of occurrence of fetal QRS complexes in the respective records. Each number is the elapsed time in milliseconds from the beginning of the record to a fetal QRS. The .csv files contain five columns; the first two lines describe the contents of these columns and the units. Beginning on the third line, the first column contains the elapsed time in seconds from the beginning of the record to the time of observation of the samples of the four signals in the remaining columns. So, for example, the first number in a01.fqrs.txt is 355, and this marks a fetal QRS that occurs 355 milliseconds after the beginning of record a01. The corresponding samples are those in the line of a0.csv that begins with 0.355 (which is the 358th line in a0.csv, since the first two lines are column headers and the third line begins with 0.000).
The .fqrs files contain the same information as the .fqrs.txt files, but in the binary format that is used throughout PhysioBank for annotation files. It is readable using software in the WFDB software package, but not using MS Excel. Similarly, the .dat files contain the same information as the .csv files, but in a binary PhysioBank- and WFDB-compatible format that cannot be read using MS Excel.
The special value '-' appears in the .csv files, and the special value -32768 appears in the .dat files, when the output of the A/D converter is invalid (for example, if the analog signal is out of the input range of the ADC, or if the transducer or cables are disconnected). Samples with this special value indicate that no valid observation of the signal was recorded during that sampling interval; you may either ignore them or use the information that the signal was lost as a contribution to an assessment of the quality of the signal.
It's likely that rankings for events 1 and 2 (or 4 and 5) will be correlated, but it is quite possible to obtain good results in event 1 and poor results in event 2 or vice versa.
An entry that detects fQRSs but does not determine their locations accurately will score well on event 1 and poorly on event 2. This might happen if the numbers of false negatives (missed fQRSs) and false positives (extra detections) are nearly equal, as would be expected of an optimized fQRS detector in the presence of significant noise; the FHR would be fairly accurate but the FRR intervals would not be.
An entry that includes a high-specificity fQRS detector may make many more false negative errors than false positive errors in the presence of significant noise; it will underestimate FHR and thus will score poorly on event 1, but if it is able to determine accurate locations for consecutive true positive detections, it will score well on event 2.
Participants in the open-source events (1, 2, and 3) must supply the sources (i.e., the actual source code, not just a reference to it) for their entries. We will publish a selection of these entries on PhysioNet after the conclusion of the challenge, with full credit to their authors.
Sorry, no. The open-source events require complete open-source entries, because our aim is to advance the state of the art by encouraging the development and publication of software that others can use, adapt to their needs, and develop further if they wish.
Participants are free to make use of open-source implementations of functions in open-source challenge entries. Open-source implementations of common functions such as cross-correlation are easy to find. Open-source QRS detection functions (suitable for finding the maternal QRS complexes) are also available (for example, see gqrs, sqrs, or wqrs in the WFDB software package).
Each participant-team may submit up to 10 entries, and they can be any mixture of open-source and non-open-source entries.
Note: before running this new tool please make sure that the directory in which you are running it on has been properly and fully backed up. The scoring procedure requires a JAR file, that can be obtained by downloading and installing the WFDB App Toolbox for MATLAB.The scoring requires that all the data and annotations for one set be stored in the same directory. For instance, in scoring record a01, the following files are necessary at the testing directory:
NON-MATLAB Users can score their entries by running the following command from a terminal or command prompt (assuming the WFDB App toobox was installed at /home/foo/wfdb and the annotation data reside in /home/foo/pn2013/data):
java -cp "/home/foo/wfdb/mcode/wfdb-app-JVM6-0-0-2.jar" org.physionet.wfdb.Score2013 a01 /home/foo/pn2013/data/ fqrs entry1
This will score your entry using Java's virtual machine.
MATLAB Users can score their individual entries by running the following command from the MATLAB prompt (assuming same installation as the example above ):
cd /home/foo/pn2013/data; [score1,score1]=score2013('a01','fqrs','entry1')
Note for Windows users: The scoring procedure requires Administrator privileges in order to generate some temporary cache files. To run the scoring procedure in MATLAB or at the command prompt, before you start the MATLAB or command prompt make sure you right click on it and select the "run as Administrator" option. The source code for the patchtann.c, the JAR file, and the toolbox can be found here and you can compile it from source if you wish.
The challenge organizers wish to thank Joachim Behar, Gari Clifford, Marcelino Martinez, Dawid Roj, Reza Sameni, Anton Tokarev, and their colleagues for their generous contributions of data and expertise in support of this challenge.
Kennedy RG. Electronic fetal heart rate monitoring: retrospective reflections on a twentieth-century technology. J R Soc Med 1998 May; 91(5): 244–250.
Sameni R, Clifford GD. A review of fetal ECG signal processing; issues and promising directions. Open Pacing Electrophysiol Ther J. 2010 Jan; 3:4–20.