Prenatal risk assessment of Down syndrome by probabilistic classifiers

Abstract

Over the last 20 years, new technology has improved the methods of detection of fetal abnormalities, including Down syndrome. While there are ways to diagnose Down syndrome by obtaining fetal tissue samples by amniocentesis or chorionic villus sampling, it would not be appropriate to examine every pregnancy this way. Besides greatly increasing the cost of medical care, these methods do carry a slight amount of risk to the fetus. So non-invasive methods such as characteristics and screening analysis have been developed to try to identify those pregnancies at "high risk". These pregnancies are then candidates for further diagnostic testing. In this thesis, we address the decision-making problems in diagnosing Down syndrome cases from the machine learning perspective aiming to decrease invasive tests. Initially, we present a comprehensive and comparative analysis of the classification techniques in Down syndrome prediction. In parallel, we evaluate the predictor effects of input features in order to eliminate the redundant features and decide the optimum feature subset leading to the highest prediction performance. Later, we focus on improving the classification performance either by parameter optimization or by improving the information content of the data. First we handle the problem of imbalanced class distribution. As a solution to imbalance class problem we analyse decision threshold optimization and re-sampling the training data techniques. Secondly, we use probabilistic classifiers based on applying Bayes Theorem, Naive Bayes and Bayesian Networks, to predict the Trisomy 21 case. In contrast to probabilistic classifiers we also apply some of widely used and well known classifiers such as Decision Tree, Support Vector Machine, Multi Layer Perceptron, and k-NN. In this thesis, we aim to evaluate the probabilistic classifiers performance with respect to these methods. This comparison is based on performance metrics such as sensitivity, specificity, accuracy and Receiver Operating Characteristics. The results of the experiments show that (i) probabilistic classifiers enable acceptable prediction of Trisomy 21 case and (ii) the classification performance can be improved by using the proposed techniques in this study.