The performance of three visual coding procedures and three computer programs in classification of electrocardiograms according to the Minnesota Code.

Abstract

A test library composed of the ECG's of 228 patients with clinically proven myocardial infarction and 294 subjects without clinical evidence of infarction was used to assess the performance of three visual coding procedures and three computer programs designed to classify ECGs according to the Minnesota Code. The results showed that visual coding performed by one experienced senior coder tended to be more consistent than visual coding relying on two less experienced coders and arbitration of disagreements by a supervisor. There was no significant difference in coding results when only one preprocessed complex was coded in comparison with the more elaborate coding of the whole source ECG using majority rule. The coding performance of the three computer programs was similar to that of the visual coding procedures. It is concluded that computer coding of ECGs according to the Minnesota Code is feasible. Combined optimal use of automated coding and visual verification of selected items may still further improve coding precision. However, when judged against an ECG independent standard, the accuracy of all coding procedures in discriminating infarcts from non-infarcts according to the Minnesota code criteria is rather limited. 'Soft' criteria give a reasonable sensitivity with low specificity whereas the use of 'hard' criteria with adequate specificity results in a substantial drop in sensitivity.