Distinguishing Ischemic from Non-Ischemic ST Changes

Introduction

For the fourth annual PhysioNet/Computers in Cardiology Challenge, we propose a provocative question of considerable clinical interest:

Is it possible to tell the difference between transient ST changes in the ECG that are due to myocardial ischemia, and those that are not?

For many years, a simple answer ("no") was considered to be the final word on this question. Myocardial ischemia results from insufficient oxygen delivery to the myocardium. To diagnose myocardial ischemia definitively, it is necessary to document that blood flow, blood oxygen saturation, or both have been compromised to an extent that the oxygen demands of the myocardium are not satisfied. These diagnostic criteria are typically established by imaging the coronary arteries. Since the ECG does not contain direct information about blood flow or oxygen saturation, it cannot be used to diagnose ischemia.

It may be possible, however, to establish inferential associations between specific features of the ECG and myocardial ischemia. One such association, between transient ischemia and changes in the ST segment of the ECG, is very widely known, and is understood to be highly sensitive, but not specific. It has long been known that repolarization of ischemic myocardial regions is abnormal, that these abnormalities are visible in the ST segment, and that they can be quantified by measuring the deviation of certain portions of the ST segment from baseline measurements. It is also known that deviations in these ST segment measurements can result from a wide variety of other causes, including changes in heart rate, conduction pattern, position of the subject, and noise in the ECG. As a result, observations of transient ST changes are considered suggestive of ischemia but are not sufficient for a definitive diagnosis, absent conclusive evidence from imaging studies.

Even in subjects who are known to have myocardial ischemia, ST changes are not considered a basis for definitive diagnosis of individual episodes of ischemia. In a subject with an old myocardial infarction, for example, the infarct may result in an ST segment with a persistent abnormal pattern (in the frame of reference of the heart). This fixed pattern appears to change with the subject's body position (upright, supine, etc.) because of movement of the ECG elecrodes relative to the heart. Thus many of those subjects who are most likely to experience ischemia are also among those most likely to have non-ischemic ST changes.

Therapeutic intervention to reduce or eliminate transient ischemic episodes can make a significant difference in quality of life for affected subjects, and may reduce mortality and morbidity in this population. Assessment of the effectiveness of therapy is substantially hindered by the lack of a reliable way of identifying ischemic episodes during activities of daily living, in which imaging studies are not possible. If it were possible to distinguish between ischemic and non-ischemic ST changes in ambulatory ECG recordings made during subjects' normal activities, the benefits would be immediate and substantial, in terms of a reduction in the time needed to determine and validate effective therapies, hence in the risk and pain experienced by the affected subjects.

This year's challenge topic encourages participants to develop novel approaches to analysis of transient ST changes using the recently-completed Long-Term ST Database, a meticulously annotated collection of 86 recordings of 2- and 3-lead long-term (20-24 hour) ECGs. Each ST change that meets criteria of clinical significance has been carefully studied by a team of expert annotators, who have drawn upon all available evidence to determine which of these events are consistent with a diagnosis of myocardial ischemia, and which are consistent with other causes. Half of these 86 recordings have been contributed to PhysioNet and are available to participants as a learning set. The remaining recordings form the test set.

Participants are challenged to design and implement algorithms that can closely mimic the decisions made by the expert annotators, classifying the ST changes (events) in the test set as ischemic or non-ischemic. The algorithms are not required to detect the events, but only to classify each given event as ischemic, non-ischemic, or indeterminate.

Organization of the Challenge

As noted above, the data used for this challenge come from the Long-Term ST Database. The learning set consists of the 43 records available from PhysioNet. Participants should train their algorithms using these records. The test set consists of the other 43 records.

To enter the challenge, participants will submit their classifiers by email to PhysioNet, where the entries will be compiled and used to classify the ST events in the test set. Each algorithm will receive a score determined by the number of correctly classified events, less the number of incorrectly classified events (those left unclassified will not affect the score). Scores will be returned to participants by email, and high scores will be posted on PhysioNet and updated throughout the challenge period. Participants may revise and resubmit their entries until the challenge deadline of noon GMT on Friday, 12 September 2003.

All participants are encouraged to submit an abstract to Computers in Cardiology 2003 describing their approach to the challenge. (When submitting your abstract, choose the topic "Computers in Cardiology Challenge".) Abstracts are due on Thursday, 1 May 2003 (note: this deadline has been extended to Thursday, 8 May 2003); details are available on the Computers in Cardiology web site. If your abstract is accepted, you will be expected to prepare a four-page manuscript (due on Tuesday, 23 September 2003) for publication in the conference proceedings, and you will have the opportunity to discuss your work at the conference. To be eligible for an award, you must submit an abstract and attend the conference.

The eligible participant whose algorithm receives the highest score will receive an award of US$1000, to be presented at Computers in Cardiology 2003 (in Thessaloniki, Greece, 21-24 September 2003). A selection of the classifiers will be posted on PhysioNet following the conference.

Developing an entry

Use the learning set to develop criteria for classifying the ST events. We recommend that you begin by copying a set of input files for one record of the learning set into an empty local directory. The files that your program will be permitted to read are:

• the header (.hea) and signal (.dat) files, permitting access to the digitized ECG signals;
• the beat annotation (.atr) and ST measurement (.16a) files, providing QRS times of occurrence for each beat, and continuously updated ST-segment measurements based on 16-beat moving averages;
• the .stf file, containing the ST level, reference, and deviation for each input signal, at two-second intervals throughout the recording
• the .klt file (decompressed from the .klt.zip files available on-line), containing time series of ST and QRS principal components

In addition, your program will need to have a copy of the .epi file for the record. These text files have been prepared for this Challenge from the .stb reference annotation files of the Long-Term ST Database; they contain the times of significant ST changes, but not the classifications of those events. Your program is not expected to detect the events, but to classify them, so this file is available to substitute for an ST change detector.

Your program may use any or all of these files as a basis for classifying the ST events. In principle, all of the other files are derivable from the signal (.dat) file, but you are not expected to do so! The other files are provided for use as shortcuts to a solution of the challenge problem; in a clinical application, it would be necessary to integrate the code needed to detect and classify the QRS complexes, measure the ST deviations, and detect the ST events.

For example, to work with record s30701 of the learning set, download

• s30701.hea
• s30701.dat
• s30701.atr
• s30701.16a
• s30701.stf
• s30701.klt.zip (decompress this using unzip, pkzip, winzip or a similar utility)
• s30701.epi

  1 40125 2 ?
  2 40129 1 ?
  3 64361 0 ?
  4 64361 1 ?
  5 76639 2 ?
  6 76647 1 ?
  7 77171 2 ?
  8 77551 2 ?
  9 79967 1 ?
 10 79975 2 ?

At the end of each run, your program must have copied the .epi file into a .epo file, replacing the '?' placeholders with its classifications. Use 'I' to mark ischemic and 'N' to mark non-ischemic events; you may leave indeterminate events marked with '?'. For example, the correct classifications for record s30701 are:

  1 40125 2 N
  2 40129 1 N
  3 64361 0 N
  4 64361 1 N
  5 76639 2 I
  6 76647 1 I
  7 77171 2 I
  8 77551 2 I
  9 79967 1 I
 10 79975 2 I

Scoring

Your program's score will be determined by comparing the output .epo files with a set of reference .epr files. The .epr files are identical to the .epi files, except that the classification of each ST event, based on the .stb annotations, is included in place of the '?' markers. A point is added to your score for each match (I/I or N/N), and a point is deducted for each mismatch (I/N or N/I). ST events left unclassified ('?') do not affect your score. (A set of .epr files for the learning set is available here for use while you are developing your classifier. The .epr files for the test set are not available; don't ask!)

The number of events per record varies considerably, from fewer than ten to several hundred. To avoid giving undue weight in the score to the handful of records that have a majority of the events, the .epr files for the test set contain no more than 20 events each (which have been chosen at random from all of the events in those records with more than 20 annotated events). Only these events will be used as the basis for scoring the entries; the others will not be counted. The same set of .epr files will be used to score all entries.

How to enter

1. Begin by downloading stclass.c (to be used unmodified) and analyze.c (to be used as a template for your entry).
2. You will need to write functions in standard (ANSI/ISO) C to replace the initialize, analyze, and finalize functions in analyze.c:
   • initialize is called once, before analyze is called the first time, and before any of the input files have been opened. Use this function to set up any variables needed by your classifier.
   • analyze is called once per ST event. Its inputs are an event ID number, which starts at 1 for the first ST event in each record and is incremented by 1 for each subsequent event; the time of the event (elapsed time in seconds); and the signal number (0, 1, or 2) of the affected signal. In each run, all of the ST changes identified by the expert annotators within a single record will be presented to the analyze function in time order.
   • finalize is called once, after all of the input files been closed. Your algorithm's classifications are recorded immediately after finalize exits.
3. Use label (defined in stclass.c) to record your classifications.
   • label accepts two input arguments. The first is an event ID number, and the second is the label you wish to assign to that event ('I', 'N', or '?'). You can invoke label to mark any of the events at any time (so, for example, you can invoke label from your analyze function if you wish to label the events one at a time, or you can invoke label once per event from your finalize function after accumulating information about all of the events. Your algorithm can relabel any event by invoking label a second (or third, ...) time with the same event ID. Any events that you do not label are marked as '?'.
4. You may, if necessary:
   • define additional functions
   • define global variables
   • include other ANSI/ISO C standard header (.h) files
   • allocate memory as needed using malloc or similar functions
   • invoke other functions from the ANSI/ISO C standard library and math library, and from the WFDB library
   • write to the standard output or the standard error output (for debugging purposes)
   • create temporary files in the current directory

   If you create temporary files, do so within the current directory only, and use file names beginning with temp. Any files created will be removed between runs (you cannot save information from one run to use in another).

5. You may not:
   • modify main or any other part of stclass.c
   • use chdir or any other means to change the current directory
   • invoke fork, system, or any of the exec family of functions to start another program or another process
   • incorporate code or data that cannot be made freely available after the conclusion of the Challenge
6. All code will be reviewed before being compiled or run. Please keep your code neat. If we can't figure out what your program does, we won't run it!
7. All code must compile cleanly using:

                gcc -Wall stclass.c -lm -lwfdb
   

   There must be no errors or warnings of any kind. See the WFDB Software Package for information about the WFDB library.

   If your program does not make use of the .dat, .atr or .16a files, it can be compiled without the WFDB library using:

   	     gcc -Wall -DNOWFDB stclass.c -lm
   

8. Your program must run to completion within a reasonable time. A reasonable time is 5 minutes or less for a 24-hour record running on a 1 GHz Athlon under Linux; we will not disqualify programs that slightly exceed this limit.
9. Test your entry before submitting it. Don't forget to include your name, affiliation, and email address in the comment block at the top of analyze.c. Once you are ready, send a copy of your version of analyze.c (source only; do not send binaries) via email to webmaster@physionet.org with a subject line of analyze.c. Please send analyze.c as plain text, not as HTML or as a word-processor formatted attachment.
10. You will receive an email confirmation of your entry once it has been reviewed. If an entry fails to meet any of the requirements for a valid entry, the email will indicate in general terms the nature of the problem (e.g., compilation error), but you are responsible for debugging your entry. Each valid entry will be assigned an entry number, which will be indicated in the email confirmation.

Once your entry has been given a number, we will run it on the test set and you will receive a score by return email. The top scores will be posted on PhysioNet and will be updated as new entries arrive. You may revise and resubmit your entry if you wish; note, however, that the challenge organizers will give priority to new participants, so that there may be a delay in receiving scores for revised entries.

We will continue to accept entries until noon GMT on Friday, 12 September 2003. All valid entries submitted before this deadline will be scored.

At Computers in Cardiology 2003 (in Thessaloniki, Greece, 21-24 September 2003), a prize of US$1000 will be awarded to the top-scoring eligible participant. Immediately following the conference, a selection of the programs entered will be posted with full credit to their authors, and they will be made freely available under the GPL (or another open source license of the author's choice).

Members and affiliates of our research groups at MIT, Boston University, Harvard Medical School, Beth Israel Deaconess Medical Center, and McGill University are not eligible for awards, although all are welcome to participate.

To qualify for an award, a participant must do all of the following:

1. Submit an abstract describing his or her classifier to Computers in Cardiology, no later than Thursday, 1 May 2003 (note: this deadline has been extended to Thursday, 8 May 2003).
2. Submit a valid entry no later than noon GMT on Friday, 12 September 2003.
3. Attend Computers in Cardiology 2003, 21-24 September 2003, in Thessaloniki, Greece.

Important dates

All deadlines are at noon GMT unless otherwise indicated. Late submissions will not be accepted.

Thursday, 8 May 2003

Deadline for submission of abstracts for Computers in Cardiology 2003.

Friday, 12 September 2003

Deadline for submission of entries to PhysioNet.

Sunday-Wednesday, 21-24 September 2003

Computers in Cardiology, Thessaloniki, Greece.

Where are the .epi and .epr files?

These files are not part of the Long-Term ST Database; they were created specifically for this challenge, and they can be found here [epi-files] and here [epr-files].

Are the classifiers allowed to read the clinical notes included in the .hea files?

Parsing those notes to understand how to classify individual episodes would be an impressive accomplishment ... but that is not the intent of the challenge! The .hea files made available to submitted classifiers for the test set will be stripped of these notes.

Why don't you have a challenge about ...?

Each year, we receive many suggestions for challenge topics. We encourage you to write to us with further suggestions.