Automated Classification of Deceleration Patterns in Fetal Heart Rate Signal using Neural Networks

Abstract

Correct classification of deceleration patterns in fetal heart rate signal is crucial issue for determining the fetal intrauterine distress of the fetus. Deceleration patterns lasting less than two minutes are divided into two classes: episodic decelerations and periodic ones. Periodic patterns are characterized by correlation with uterine contraction, while episodic decelerations do not show such relation. The research material includes 101 cardiotocographic records (total time 285 hours) from which, the clinical experts selected 383 patterns for further classification. Nineteen different parameters of quantitative description of deceleration were used as the input variables for the neural networks (NN) classification system. It turned out that there was a group of 11 parameters which can be removed because they have very weak influence on the classification process. Quality indices of the developed neural networks (from 93 % to 99 %) and the ROC curve indexes (from 0.9863 to 0.9944) explicitly show that the proposed NN structures are very efficient for the classification of deceleration in fetal heart rate signal.