Ensemble based technique for the assessment of fetal health using cardiotocograph – a case study with standard feature reduction techniques

Intrauterine fetal hypoxia is one of the leading cause of perinatal mortality and morbidity. This can eventually lead to severe neurological damage like cerebral palsy and in extreme cases to fetal demise. It is thus necessary to monitor the fetus during intrapartum and antepartum period. Cardiotocograph (CTG) as a method of assessing the status of the fetus had been in use for last six decades. Nowadays it is the most widely used non-invasive technique for the continuous monitoring of the fetal heart rate (FHR) and the uterine contraction pressure (UCP). Though its introduction limited the birth related problems, the accuracy of interpretation was hindered by quite a few factors. Different guidelines that are provided for the interpretation are based on crisp logic which fails to capture the inherent uncertainty present in the medical diagnosis. Misinterpretations had led to inaccurate diagnosis which resulted in many medico-legal litigations. The vagueness present in the physician’s evaluation is best modeled using soft-computing based techniques. In this paper authors used the CTG dataset from UCI Irvine Machine Learning Data Repository which contains 2126 data and each data-point is represented by 37 features. Dimensionality of the feature set was reduced using different automated methods as well as manually by the physicians. The resulting data sets were classified using various machine learning algorithms. Aim of this study is to establish which set of features is best suited to give good insight into the status of the fetus and also determine the most effective machine learning technique for this purpose. The accuracy of the outcomes were measured using statistical methods such as sensitivity, specificity, precision, F-Measure, confusion matrix and kappa value. We obtained an accuracy of 99.91% and kappa measure of 0.997 when the feature set was reduced using MRMR.