A new ECG classification system for myocardial infarction based on receiver operating characteristic curve analysis and information theory.

Abstract

SUMMARY An optimized three-lead ECG hierarchial decision-tree type of classification system for myocpardial infarction is presented. For selection of the best threshold values for each criterion and the best association of features, we developed a procedure based on "receiver operating characteristic" (ROC) curve data analysis and information theory. Optimization was obtained through maximization of information content of the criteria. The classifier is based on nine measurements that can be easily obtained by hand (Q X duration, Q/R Y amplitude, R Y amplitude, Q/R Y duration, Q Z amplitude, QRS and Taxes in the horizonital plane, Q Z duration and R Z amplitude) and achieved a satisfactory performance in an independent group of patients (true-positive ratio 0.853, false-positive ratio 0.105, average information content 0.308 bits).THE ECG variables most critically affected by myocardial necrosis are well established. However, the best threshold values for these variables and the best association of features for ECG classification are still controversial.Quantitative electrocardiography has revealed that the performance of an ECG test may be changed at will by altering the cutoff values of the respective variables. Selection of any one of these values may be governed by a specific objective, such as the identification of as many patients with a given condition as possible or, conversely, the classification as abnormal of the least number of patients without the disease. Usually, a reasonable compromise between these two extremes is adopted.When the values attributed to a given ECG variable are gradually changed in successive classification experiments and the corresponding true-positive and false-positive ratios (TPR and FPR) are plotted against each other, the result is the so-called received operating characteristic (ROC) curve. These plots, developed as a method of observer performance analysis in detection experiments of electromagnetic signals transmitted through noisy channels,' are becoming of incre'asing interest in the study of medical decisionmaking problems, namely, the evaluation of diagnostic tests.24Clinical diagnosis is essentially a classification procedure. The fact that different conditions to be distinguished or classified share a number of features lends a degree of uncertainty to any clinical diagnosis. Clinical tests, such as the ECG, aim at reducing the degree of this uncertainty. Information theory enables the measurement of uncertainty before and after a test is undertaken. The information content or clinical effectiveness of the test can then be computed on the basis of this measurement.7 8A procedure for selection of the best threshold val- ues of ECG variables, as well as the best association of features for classification purposes, which is based on maximization of information content, is being used in our laboratory. The purpose of this report is to present the results obtained with this procedure for developing a decision-tree classification system for myocardial infarction (M...

Interpretation of Short ECGs with a Personal Computer: The Porto Program

“…This Code was mainly designed for epidemiological investigations, but its authors consider it adequate for clinical use, provided the threshold criteria with the highest level of specificity are used [3], as we did. Although the diagnostic accuracy of our ECG computer system, as evaluated against independent clinical evidence, is reasonable, later studies of our group suggest that it could be enhanced by means of criteria optimized by a new and powerful method reported elsewhere [5,12,13].…”

Evaluation of GDx parameters by using information theory

“…The clinical effectiveness of a test depends on its information content, which is related to the uncertainty remaining after the test. At this point, information theory seems as powerful a tool for handling such problems (4,16).…”