Procedures for improvement of quality of diagnostic advisor built on binary classifiers

Abstract

The work aims to improve quality of classification of diagnostic advisor on example of problem of differential diagnosis of mild forms of hemostasis, such as von Willebrand's disease, coagulopathy, deaggregating thrombocytopathy, combined pathology of hemostasis. To solve problems we have compared efficiency of diagnostic advisor, developed by relaxation iterative algorithm GMDH according to two approaches. In first variant, system was developed on binary classifiers, built on principle of one against all; in second diagnostic problem was solved by a single classifier for four classes. For this task advantages of first variant were shown. The further step was improvement of system by introduction of procedures for solving classifiers conflicts. With this purpose we proposed to design additional functions of classification on data sets with different combinations of samples of classes that must be differentiated: classification of each pair of diagnoses, classification of diagnosis from a couple of other diagnoses, classification of diagnoses in pairs. Thus, a multi-classification system is formed, where at each further level conflicts left unsolved at lower levels, are solved. This approach should be applied while formulating a criterion of diagnostic system as the maximum number of correct diagnosis