Novel approach for fetal heart rate classification introducing grammatical evolution

Abstract

Fetal heart rate (FHR) variations reflect the level of oxygenation and blood pressure of the fetus. Electronic Fetal Monitoring (EFM), the continuous monitoring of the FHR, was introduced into clinical practice in the late 1960s and since then it has been considered as an indispensable tool for fetal surveillance. However, EFM evaluation and its merit is still an open field of controversy, mainly because it is not consistently reproducible and effective. In this work, we present a novel method based on grammatical evolution to discriminate acidemic from normal fetuses, utilizing features extracted from the FHR signal during the minutes immediately preceding delivery. The proposed method identifies linear and nonlinear correlations among the originally extracted features and creates/constructs a set of new ones, which, in turn, feed a nonlinear classifier. The classifier, which also uses a hybrid method for training, along with the constructed features was tested using a set of real data achieving an overall performance of 90% (specificity = sensitivity = 90%).

Introduction

Antepartum and intrapartum fetal surveillance is widely based on monitoring, analysis and evaluation of fetal heart rate (FHR). FHR is acquired by cardiotocographs along with the uterine activity (UA) and both signals are printed on a single strip of paper termed cardiotocogram (CTG). FHR is measured in beats/minute and it is acquired either by scalp electrodes, after the rupture of the membranes, or by an external sensor attached on the mother's abdomen [1]. Fig. 1 shows a typical CTG. FHR signal alterations are used to evaluate the fetal health condition, so as to early diagnose fetal stress and distress. In the latter case, the obstetrician has to intervene to prevent potential compromise and irreversible damage, such as cerebral palsy or death. Especially during the crucial period of labor, FHR monitoring is used as the main screening test of the fetal acid base balance [2].

FHR monitoring has been used during the last four decades as the main electronic fetal surveillance test. However, extensive studies have shown significant inter-observer and intra-observer variation in FHR analysis and interpretation [3], [4], [5]. These inconsistencies in interpretation and the increase of false positive diagnosis have created skepticism. On the other hand, the advances in pattern recognition methods, along with new signal processing techniques, have paved the way towards automated approaches. Therefore, many researchers proposed computer-based systems, in an attempt to monitor and evaluate the condition of the fetus in a reliable, reproducible and effective way [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23], [24], [25], [26], [27], [28], [29], [30], [31].

The paths followed by researchers in the field are quite diverse. Mantel et al. [8], [9], as well as Taylor et al. [11], used an iterative procedure to extract the main morphological features of FHR. Bernardes et al. [17], [18], [19] developed an automatic method for the analysis and recognition of CTG and subsequent extraction of morphological features. Magenes et al. [20], [21] and Kol et al. [22] employed artificial neural networks for the interpretation of FHR recordings. Chung et al. [23] developed an algorithm to analyze and predict fetal acidosis. Salamalekis et al. [24], Struzik and Wijngaarden [25] and Georgoulas et al. [26] employed wavelets for the analysis of the FHR signal. Ifeachor et al. developed an expert system [27], which they subsequently upgraded into a fuzzy system [28], for the interpretation of CTG records. Alonso-Betanzos et al. merged FHR and contextual analysis of all pathological and physiological aspects involved in fetal monitoring. They developed at first an expert system called NST-EXPERT [29], [30], which later evolved to create computer aided fetal evaluator (CAFE) [31].

Based on the belief that the FHR may convey much more information than what is usually interpreted by obstetricians [6], we investigate and propose an advanced methodology to analyze and interpret FHR. This method is used to early diagnose fetal acidemia. This innovative method utilizes a novel approach for feature construction based on the extracted features from the FHR signal. The constructed features are fed to a multilayer perceptron (MLP) nonlinear classifier [32], [33], with very promising results.

The original FHR features are derived from three domains: the time domain, the frequency domain and the morphological domain. The last one utilizes medical definitions of “morphological” features, which have already been used with quite a success both in the antepartum [20], [21] and intrapartum case [34], [35].

The recently proposed approach of grammatical evolution is applied to construct new artificial features from the actual ones. Grammatical evolution [36] is an evolutionary methodology that, as in the case of genetic programming [37], can evolve complete programs. In our case, the evolved programs are mathematical expressions/functions of the originally extracted features. These constructed artificial features are then used to classify the FHR signal.

This paper is structured as follows: Section 2 gives a brief introduction to the grammatical evolution method. Section 3 presents the overall proposed procedure and the stages preceding the construction phase. Section 4 describes the implementation of the grammatical evolution for feature construction while Section 5 presents a brief analysis of the new features in terms of their “quality”. Section 6 compares the experimental results for different implementations of the proposed scheme and, finally, Section VII concludes the paper and ideas and directions for future work are discussed.