Composition of Decision Trees for Severity of Chronic Obstructive Pulmonary Desease Recognitio
Keywords:
Decision Trees, Compositional Classifiers, Medical Diagnostics, COPDAbstract
Purpose: Chronic obstructive pulmonary disease is one of the most prevalent pulmonary diseases, and spirometry is one of the
most important methods to diagnose its severity. Unfortunately, spirometry is not widely available in Russian hospitals and clinics.
This paper proposes an algorithm of COPD severity diagnostics without spirometry. Methods: As a mathematical framework for the
diagnostics, decision trees were chosen. On their base, a two-level compositional classifier was implemented. The primary decision
tree provides a preliminary diagnosis refined in the second step by another more specialized decision tree. Results: The low accuracy
of the classifier can be improved if two conditions are met: the confusion matrix has block-diagonal structure, and the classifiers
built for each block have a higher accuracy than the original classifier. In order to improve the cross-validation accuracy of the
classifier from less than 55%, a two-level classifier scheme is proposed and tested. First-level classifier is refined by a number of
secondary classifiers built for the diagonal blocks of the original confusion matrix. The proposed solution improves the accuracy of
the COPD severity diagnostics from 52,5 to 65%. Practical relevance: The differential diagnostics of COPD severity can be performed
with satisfactory accuracy even in hospitals without spirometry equipment. The proposed method for improving the classifier
accuracy can be applied in other diagnostics classifiers, if a set of solvers more competent in narrow areas than the original ones
is successfully built.