Empirical Classification Accuracy Assessment of Various Classifiers for Clinical Diagnosis Datasets

Abstract

Classification is a predictive data mining task. Nowadays, it is also playing a pivotal role in the field of medical diagnostic towards early disease predictions. The aim of applying different classification techniques in diseases like cancer, diabetes, kidney infections, etc., is not to undermine the decision of doctor, but the outcomes determined from the classifiers might augment the correct treatment initiatives. The classifiers developed for medical diagnosis should be validated on reliable results to be trustworthy by doctors. In this research work, the authors attempted to assess the classification accuracy of different classifiers on datasets taken from UCI with cross-validation. Majorly, SVM, logistic regression, ML perceptron, Naïve Bayes, fuzzy logic, k-nearest neighbours, random forest, and J48 are used for experimentation purposes. The performance measures like accuracy, RO curve, kappa statistics, MAE, RMSE, and model building time are used on WEKA. The authors have chosen datasets specifically related to liver, heart, and diabetes among widely spread most life-threatening diseases. Experimental results show that random forest demonstrated the best classification and prediction capability over other classifiers and chosen datasets.