Journal of electrocardiology vol:20 issue:2 pages:83-92
The performance of logistic (LOG) and linear discriminant analysis (LDA) has been studied, both for the conventional 12-lead electrocardiogram (ECG) and the orthogonal Frank 3-lead electrocardiogram (VCG), using a large validated data base. Classification rules were derived form a learning set (N = 2446) and applied to a test set (N = 820) to differentiate between normal, left, right and biventricular hypertrophy, anterior, inferior and combined myocardial infarction (MI). Total accuracy of LOG, assuming no normal distribution and using population proportions as prior probabilities, was up to 3% higher than that of LDA, depending on the number of variables used. The 12- and 3-lead LOG and LDA formulas resulted in very similar accuracy rates, i.e., between 67 and 70% for the seven-group and between 77 and 84% for the five-group analysis. LDA posterior probabilities were systematically more extreme than LOG ones. Correct classification of normals' specificity by LDA was 5 to 9% higher, but sensitivity for different groups was 1.5 to 10% lower than by LOG, with sample size proportions as priors. Specificity could be improved by changing the priors at the cost of lower sensitivity and vice versa, both for the LDA and LOG models. Classification results at 95% specificity were only slightly different, except for anterior MI where LOG scored 6% better. Other measures of performance demonstrated that the LDA model was overconfident and that the LOG model fitted better the real class membership of the patients. In conclusion, logistic ECG and VCG models improve the total accuracy of classification by about 1 to 3% when compared to LDA. More importantly, reliability of classification represents the improvement we want to emphasize. These methods may enhance the diagnostic utility of the ECG and VCG in routine practice.