Using classification and regression tree to detect hematology abnormalities

The detection of the abnormal blood cells and particles in a blood test is essential in medical diagnosis. The detection rules, which are usually implemented in the widely used automated hematology analyzer, are therefore critical for the health and even lives of millions of people. The research endeavor of this thesis is on generating such detection rules using a supervised machine learning algorithm. The first part of this thesis studies the hematology data and surveys the popular classification algorithms. In the second part, the selected algorithm, CART, is implemented with deliberately selected parameters. In the third part, a modification of the algorithm, logical pruning with Enclose the Normal principle, is exercised. To extend the algorithm and to achieve better performance, I developed and implemented the idea of decision tree combinations. The research has proven to be successful by the achievement of good performance and reasonable detection rules.