The stability of decision-Theoretic electrocardiographic classifiers based on the use of discretized features

The stability of electrocardiographic classification of myocardial infarction was evaluated using linear and quadratic discriminant functions and the minimum euclidean distance to means classifier. Each classifier used features discretized at binary and ternary levels and the changes in classification accuracy were determined when the value of each discretized feature was altered systematically or randomly. The data file was composed of ECGS of 128 normal subjects and 237 subjects with old myocardial infarction documented by ECG independent clinical evidence. The quadratic discriminant function is found to be quite unstable in comparison with the linear discriminant function, which in turn is slightly less stable than the minimum euclidean distance to means classifier. Binary and ternary features yield nearly equally stable results. With these decision-theoretic classifiers using discretized ECG features the fall in classification accuracy as a result of wrong representation of one feature rarely exceeds 20% and is usually less than 10%.