HYBRID DECISION TREE CLASSIFICATION FOR MEDICAL DIAGNOSIS

Abstract

newlineMedical decision support systems (MDSS) are intelligent computerized platforms designed to interpret complex medical information and to generate the foundation for reliable decision making in medical practice. MDSS is expected to improve the quality of medical care by assisting healthcare professionals with wide range of clinical decisions. newlineIn the process of development of such systems, many factors have been attributed but inadequate information has been identified as a major challenge. To reduce the diagnosis time and improve the diagnosis accuracy, it has become more of a demanding issue to develop reliable and powerful MDSS. The medical diagnosis by nature is a complex and fuzzy cognitive process, hence soft computing methods, such as fuzzy decision tree classifiers have shown great potential to be applied in the development of MDSS. newlineThis thesis discusses hybrid decision tree (DT) classification approaches for computer aided medical diagnosis. The first among these is a new fuzzy decision tree (FDT) classifier named Fuzzy_HSM, which is used for soft decision making. The next approach deals with the biased decision making problem of DT classifiers which comes due to information theory based node split measures used for decision tree induction. The third approach is an intelligent two phase model designed for computer-aided detection and diagnosis (CADe/CADx) of medical data. Significant attributes with their significance weights (SW) are computed in Phase-I (acts as CADe) and a decision tree based clustering approach is used in Phase-II (acts as CADx) for medical data classification. An ant colony optimization (ACO) based fuzzy rule miner named ANT_FDCSM is proposed in the fourth approach. The basic idea is to generate an optimal and comprehensible rule set for medical diagnosis while balancing accuracy and sensitivity count. The performance analysis of proposed approaches is done on several general and medical domain data sets taken from UCI repository. A significant performance gain is found across all approaches which demonstrate the superiority of proposed approaches over well-known approaches proposed in the literature. newline