Computerized electrocardiographic interpretation: an analysis of clinical utility in 5110 electrocardiograms

This study describes the initial experience with a computerized electrocardiographic interpretation system in a teaching hospital. The sensitivity and specificity of the 13,375 diagnostic statements that were used to describe the first 5110 electrocardiograms were analysed to determine the predictive accuracy of computerized electrocardiographic interpretation. Reviewing cardiologists inserted 1320 statements, deleted 1792 statements and modified 484 computerized statements. The over-all sensitivity and specificity (and standard error [SE]) of computerized diagnosis was 90.1% +/- 0.3% and 89.6% +/- 0.2%, respectively, with an over-all positive predictive accuracy (+/- SE) of 87.1% +/- 0.3% and a negative predictive accuracy (+/- SE) of 92.2% +/- 0.2%. Sensitivity and specificity were lowest for the category of ST-T wave changes (83.1% +/- 0.8% and 84.1% +/- 0.7%, respectively) and were highest for the category of sinus rhythm (96.6% +/- 0.3% and 97.0% +/- 0.6%, respectively). The positive predictive accuracy of computerized diagnosis was lowest for the category of hypertrophies (74.2% +/- 1.0%) and was highest for the category of sinus rhythm (99.5% +/- 0.1%), while for the category of myocardial infarctions it was 87.6% +/- 0.8%. The negative predictive accuracy ranged from 96.7% +/- 0.3% for the category of hypertrophies to 81.8% +/- 1.3% for the category of sinus rhythm. We conclude that the computerized analysis of electrocardiograms has a satisfactory predictive accuracy when used in an environment with a high prevalence of abnormalities. Electrocardiograms that are classified as normal by computerized analysis may not require checking; however, all electrocardiograms with abnormalities should be interpreted by a competent electrocardiographer.