Refining the rule base of fuzzy classifier to support the evaluation of fetal condition

Abstract

The paper proposes a method to simplify a rule base of zero order Takagi–Sugeno–Kang fuzzy classifier, involving the determination of the ɛ-similar rules based on fuzzy clustering with ɛ-hyperballs. The rule simplification process is based on the concept of ɛ-insensitivity areas underlying the partitioning process of rule centers (centers of membership functions in the rule premises), which directly corresponds to the idea of rule ɛ-similarity. Clustering parameters leading to the best performance of the modified rule base, including the degree of rule ɛ-similarity, are determined by means of the evolution strategy. Since our main objective was to maintain the high performance of the resulting classifier, two rule-based simplification procedures, both called rule base refinement, are proposed. The work focuses mainly on the practical application to support the diagnosis of fetal condition based on the analysis of CardioTocoGraphic (CTG) signals. The publicly available collection of CTG recordings (CTU-UHB) was used in order to verify the quality of the introduced solutions. The classification performance was assessed with respect to the reference evaluation of fetal state determined on the basis of a retrospective analysis using the newborn outcome defined with different thresholds of the blood pH from the umbilical artery. The experiments confirmed the high generalization ability of the refined fuzzy classifier, in particular its high efficiency in supporting the qualitative assessment of fetal condition based on the analysis of parameters quantitatively describing fetal signals.