This study investigated interobserver (two observers) and intrasubject (two measurements) reproducibility of QT dispersion from abnormal electrocardiograms in patients with previous myocardial infarction, and compared a user-interactive with an automatic measurement system. Standard 12-lead electrocardiograms, recorded at 25 mm.s-1, were randomly chosen from 70 patients following myocardial infarction. These were scanned into a personal computer, and specially designed software skeletonized and joined each image. The images were then available for user-interactive (mouse and computer screen), or automatic measurements using a specially designed algorithm. For all methods reproducibility of the RR interval was excellent (mean absolute errors 3-4 ms, relative errors 0.3-0.5%). Reproducibility of the mean QT interval was good; intrasubject error was 6 ms (relative error 1.4%), interobserver error was 7 ms (1.8%), and observers' vs automatic measurement errors were 10 and 11 ms (2.5, 2.8%). However QTc dispersion measurements had large errors for all methods; intrasubject error was 12 ms (17.3%), interobserver error was 15 ms (22.1%), and observers' vs automatic measurement were errors 30 and 28 ms (35.4, 31.9%). QT dispersion measurements rely on the most difficult to measure QT intervals, resulting in a problem of reproducibility. Any automatic system must not only recognize common T wave morphologies, but also these more difficult T waves, if it is to be useful for measuring QT dispersion. The poor reproducibility of QT dispersion limits its role as a useful clinical tool, particularly as a predictor of events.