The performance of ECG classifiers using Bernoulli variables was evaluated in comparison with the same classifiers using an identical set of ECG features measured in continuum. A Bernoulli variable is permitted to take only two states depending on whether its value is below or above a given threshold. A new method was used for discretization threshold optimization. The branch-and-bound algorithm, the sequential forward selection and backward rejection procedures were used for selecting the best subsets of Bernoulli and continuous features out of the primary set of features measured from the conventional 12-lead and the orthogonal Frank-lead ECG. A linear discriminant function was used to estimate the classification accuracy of Bernoulli and continuous features in a test library composed of ECGs of 237 Subjects with old myocardial infarcts and 299 subjects without infarction. The results indicate that while Bernoulli variables perform as well as or better than continuous variables over a wide range of sensitivity and specificity, continuous features appear to yield a higher diagnostic accuracy when a high level of specificity is required. In comparison with the diagnostic accuracy of the Minnesota Code's criteria for myocardial infarction, decision-theoretic classifiers using continuous features yielded an overall improvement in sensitivity from 14 to 20% at a fixed level of specificity. The information content of the conventional 12-lead ECG appeared as good as or better than that of the orthogonal Frank-lead ECG regarding differential diagnosis between myocardial infarction and non-infarction groups when identical statistical procedures were used for the design of the classifiers.