Evaluation of Classification Algorithms used in Medical Decision Support Systems

The main focus of the study is to identify the most suitable & appropriate data mining classification algorithms being used or implemented in medical decision support systems and also to analyze, evaluate the performance of the algorithms. According to the previous studies mainly six algorithms were found to be appropriate these were J48 decision tree, Naive Bayes, IBK, Support Vector Machine, ZeroR & VFI. The different datasets were chosen for assessment; these five UCI datasets were hepatitis, dermatology, heart-statlog, liver-disorders & lung-cancer dataset. The performance measures used in this study were percent of correct classifications instances or accuracy, absolute mean error, precision, recall, and F-measure values. The experiment was being conducted using the WEKA tool and the classification algorithm analysis was done using cross validation of 10 folds. The results reveals that for dataset hepatitis the accuracy measure for the classifier Support Vector Machine & VFI was found to be highest with values 85.16% correctly classified instances each, second appropriate classification algorithm Naive Bayes with 84.51% accuracy. When considering the heart-statlog dataset the accuracy measure value of Support Vector Machine was found to be best as compared with other classifiers with value 84.07% also the mean absolute error value of 0.1593 was being the lowest. The result concludes that identification of appropriate algorithm for the use in medical decision support system is a challenging task. According to the overall performance based on average and correlation it can be concluded that for the above mentioned medical datasets the most appropriate algorithms are VFI, J48 Decision Tree & Naive Bayes with accuracy values 80.52%, 80.27% and 79.79% respectively.