Abstract:
In this thesis, we address the decision-making problems in in vitro fertilization treatment from the machine learning perspective aiming to increase the clinical success rates. Initially, we present a comprehensive and comparative analysis of the classification techniques in embryo-based implantation prediction. In parallel, we evaluate the predictor effects of input features in order to eliminate the redundant variables and decide the optimum feature subset leading to the highest prediction performance. In contrast to the limited relevant literature, our preliminary experiments demonstrate the potential of machine learning classifiers as an automated decision support tool in critical decisions affecting the success of the treatment. Later, we focus on improving the classification performance either by algorithmic enhancements or by improving the information content of the data. First, we handle the problem of imbalanced class distribution and show that decision threshold optimization and re-sampling the training data produce similar results. Second, we propose a frequency based encoding technique to effiently transform categorical variables into continuous numeric values. And third, in addition to the patient and embryo characteristics, we investigate the effect of in- dividual physicians as a human factor on the pregnancy outcome. Finally, we apply Bayesian Networks to model the embryo growth process with the objective of blastocyst score prediction. We propose a novel approach to adjust the frequency estimates for parameter learning in conditional probability tables. The results of the experiments show that (i) the standard machine learning algorithms enable acceptable prediction of implantation and blastocyst score and ii) the prediction performance can be improved by using the proposed techniques in this study. From the clinical perspective, our re- sults have practical implications in reducing multiple pregnancies, preventing waste of embryos and cancelation of transfers.
