Feature Minimization for Diabetic Disorders High Performances Prediction System-based on Random Forest Tree

Abstract

Human organ failure due to high blood sugar is considered a chronic disease. Early prediction might reduce or prevent complications due to such disorders, especially with recent machine-learning improvement techniques and the availability of electronic data from different sources. The number of diabetic patients roughly increased and may reach more than 600 million by twenty years. Transforming data into valuable and helpful information is an effort for researchers to improve the performance of ML techniques. This paper applies several types of sampling to predict 1000 samples with attributes and three diabetes class types (Random Forest tree, Hoeffding tree, LWL, NB updatable, and support vector Machine). This paper focused on most parameters that affected overall prediction accuracy. ML performances have been measured depending on the accuracy and mean absolute error for several cross-validation values before Feature reduction and after feature minimization by applying feature selection methods. It shows that Gender, Age, Blood Sugar Level (HbA1c), Triglycerides (TG), and Body Mass Index (BMI) are the most impact attributes applied. It is also shown that the Random Forest tree was the best method (97.7 and 98.6 %) with and without feature minimization, respectively, but it has a higher performance by omitting some unbalanced features from the diabetic dataset. Weight minimization has also been applied to techniques like SVM to obtain a better-searching plane and a robust model. In addition, this study specifies which parameters have weight minimization with the required analysis. Also, the feature selection method was applied to gain memory and reduce time.