An initial set of beat labels was produced by a slope-sensitive QRS detector, which marked each detected event as a normal beat. Each two-hour, two-channel ECG record was printed out in full disclosure format, each page two minutes in duration, with the addition of QRS detection marks, trend plots of ST segment displacement and T-wave amplitude (measured for each beat), and boxes for checking annotation operations. For each record, two cardiologists (neither of whom was a member of the research group which had submitted the record) were given copies of the full-disclosure printout, trend plots of mean heart rate and ST-T parameters at 10-second intervals, and record-specific transparent plastic rulers for measuring time intervals and ECG signal displacements. A heart rate scale and a two-channel reference QRST complex (taken from the first 30 seconds of each record) were printed on each ruler.

Working independently, the cardiologist-annotators visually checked the computer-generated beat labels on the full-disclosure printouts and manually corrected them, and inserted annotations indicating changes in ST and T morphology, rhythm, and signal quality. Annotations from the two cardiologists were compared and the differences were resolved by a cardiologist of the coordinating group. This method assumes that the third cardiologist is able to make a more reliable judgement since he knows both sets of annotations.

Definition of ST and T episodes

The cardiologists participating in the project jointly defined and followed a set of rules for locating ST episodes and T episodes (i.e., intervals during which the ECG exhibits significant ST segment or T-wave changes). To identify and annotate an ST episode, these criteria were applied:

• ST segment deviations are measured relative to a reference waveform for each subject (usually selected from the first 30 seconds of each record). Measurements of ST segment deviation are taken 80 milliseconds after the J point if the heart rate does not exceed 120 bpm, and 60 milliseconds after the J point otherwise.
• ST episodes must contain an interval of at least 30 seconds during which the absolute value of the ST deviation is no less than 0.1 millivolt (mV).
• The beginning of each ST episode is annotated. The beginning is located by searching backward from the time at which the absolute ST deviation first exceeds 0.1 mV. The search continues until a beat is found for which the absolute ST deviation is less than 0.05 mV, and for which the absolute ST deviation is less than 0.1 mV throughout the previous 30 seconds. An ST change annotation which indicates the beginning of the episode is placed immediately following this beat.
• The peak (i.e., the greatest deviation, positive or negative) of each ST episode is annotated. An ST change annotation is placed before the beat judged to exhibit the greatest ST deviation; this annotation contains a manual measurement of the peak ST deviation.
• The end of each ST episode is annotated. The end is located by searching forward from the time at which the absolute ST deviation last exceeds 0.1 mV. The search continues until a beat is found for which the absolute ST deviation is less than 0.05 mV, and for which the absolute ST deviation is less than 0.1 mV throughout the following 30 seconds. An ST change annotation which indicates the end of the episode is placed immediately before this beat.

To identify and annotate a T episode, similar criteria were applied:

• T deviations are measured relative to the same reference waveform which is used for measuring ST deviations. The quantity A_T is defined as the amplitude of the dominant phase of the T-wave, measured relative to baseline (at the PQ junction); if the T-wave is inverted, or if the dominant phase of a biphasic T-wave is below the baseline, A_T is negative. The T deviation is defined as the difference (positive or negative) between the values of A_T for the current waveform and for the reference waveform.
• T episodes must contain an interval of at least 30 seconds during which the absolute value of the T deviation is no less than 0.2 mV.
• The beginning of each T episode is annotated. The beginning is located by searching backward from the time at which the absolute T deviation first exceeds 0.2 mV. When an interval of at least 30 seconds is found in which the absolute T deviation does not exceed 0.2 mV, the end of that interval defines the beginning of the episode. A T change annotation is placed before the first beat of the episode.
• The peak (i.e., the greatest deviation, positive or negative) of each T episode is annotated. A T change annotation is placed before the beat judged to exhibit the greatest T deviation; this annotation contains a manual measurement of the peak T deviation.
• The end of each T episode is annotated. The end is located by searching forward from the time at which the absolute T deviation last exceeds 0.2 mV. When an interval of at least 30 seconds is found in which the absolute T deviation does not exceed 0.2 mV, the beginning of that interval defines the end of the episode. A T change annotation is placed after the last beat of the episode.
• Within T episodes which contain absolute T deviations exceeding 0.4 mV, additional T change annotations are placed whenever the absolute T deviation crosses the 0.4 mV threshold value which defines extreme T deviations. These additional T change annotations indicate the beginning and end of each such interval of extreme T deviation.

These rules were applied to each of the two signals independently; for this reason, each ST and T change annotation indicates the signal to which it applies.

Each ST and T change annotation contains a text field which describes its significance. The text field contains characters which identify the episode type ('ST' or 'T'), the signal number ('0' or '1'), and the direction of the deviation ('+' or '-'; extreme T deviations are signified by '++' and '--'). The text field of an annotation which marks the beginning of an episode contains a '(' prefix. For an annotation which marks the end of an episode, there is a prefixed 'A' and an appended 3- or 4-digit decimal number which expresses the magnitude of the peak deviation in microvolts. An annotation which marks the end of an episode has a ')' appended to the end of its text field. For example, an episode of ST depression in signal 0 with a peak (absolute) deviation of 200 microvolts would be marked by three annotations, with text fields of '(ST0-', 'AST0-200', and 'ST0-)'.

In six records (e0161, e0509, e0601, e0611, e0613, and e0615), axis shifts resulting from positional change give the appearance of real ST or T changes. These axis shifts are annotated using comment annotations. The text fields of these annotations are constructed in the same way as for ST and T change annotations, except that lower-case characters are used in order to make it easier to distinguish these axis shift episodes from real ST or T change episodes. For example, an axis shift in signal 1 which gives the appearance of a peak T deviation of 350 microvolts would be marked by three annotations, with text fields of '(t1+', 'at1+350', and 't1+)'.

Annotation types

The following types of annotations appear in the European ST-T Database reference (.atr) annotation files. The Code column shows the symbols defined in ecgcodes.h, and the Mnemonic column indicates how these annotations are displayed by WFDB applications such as WAVE, WVIEW, and pschart.

Which was the reference beat in each record?

As noted above, the expert annotators were given a clear plastic template on which had been printed a reference waveform. The position of this waveform was not recorded, however, and the original plastic templates no longer exist. The only available information about the choice of reference beat is that the waveform was taken from the first 30 seconds of the record being annotated. One may assume that the waveform was typical of those within the 30-second interval, and that if the amount of noise varied significantly within the interval, the reference was one of the cleaner waveforms.

As a practical matter for evaluation of an algorithm for automated ST analysis using this database, this question need not be an issue. Some of the patients represented in the database had prior myocardial infarctions with consequent fixed ST elevation or depression. The ST annotations in this database mark transient ST changes that are superimposed on any fixed elevation or depression. The important point is that this database's annotations provide samples not of the ST level function (the difference, for any given time, between the ECG amplitudes of the nearest beat during the ST segment and at the isoelectric point), but of the ST deviation function (the difference between the ST level function measured at any given time and during the first 30 seconds of the record). Put another way, the ST level function is the sum of the fixed elevation or depression (the reference ST level) and the transient changes in ST level (the ST deviation function).

To use this database to evaluate an ST analysis algorithm, the algorithm needs to estimate the ST deviation function, a task that requires determining its own reference ST level (using any desired method; a median of its ST level measurements made during the first 30 seconds is a commonly used approach). The algorithm's ST deviation function is the difference between its ST level function and its reference ST level. See epicmp for details on how to record an algorithm's ST deviation function in an annotation file, and how to use standard software to measure how well an algorithm's ST deviation measurements match those provided with the database.