Identifying Ranked Features through Optimized Selection Procedure towards Achieving Improved Accuracy in Diabetes Detection

World Health Organization (WHO) reported in 2018, diabetes is a severe life-threatening illness, affecting 422 million people globally. Every year, the number of people affected rises significantly. For a medically significant outcome, early identification of diabetes is indeed preferred to prevent its severity. The goal of this paper is to use various data mining approaches to determine the probability of diabetes recurrence. The data for diabetes patients was obtained from the UCI machine learning repository's Pima Indians Diabetes dataset having a total of 9 attributes. Naive Bayes, Multilayer Perception, and JRip classification algorithms were used to evaluate the accuracy of their predictions on this dataset. The result of this experiment shows that using Ranker with these algorithms exhibit significant improvement in accuracy for diabetes detection. By removing certain lower-ranked features, it was possible to increase the accuracy for every model using an optimal feature selection approach. Not only do these features contribute very little, but their incorporation also causes classification systems to be misled. This helped the JRip, Multilayer Perceptron, and Naive Bayes algorithms achieve accuracy of 78.26%, 77.73%, and 77.47% respectively.