The Effect of Feature Selection on Diabetes Prediction Using Machine Learning

Diabetes arises from inadequate insulin production or ineffective response to insulin in the body, leading to long-term health complications such as cardiovascular disorders and organ dysfunction. This study aims to enhance the accuracy of diabetes prediction using machine learning techniques. Two datasets, namely the Pima Indians Diabetes and Mendeley datasets, were utilized to evaluate the performance of various machine learning classifiers. The Mendeley dataset presented the challenge of class imbalance, which was effectively addressed by employing a novel data balancing technique called Random Data Partitioning with Voting Rule (RDPVR), resulting in improved accuracy in diabetic prediction. Feature selection methods, including Recursive Feature Elimination (RFE), Analysis of Variance (ANOVA), and Step Forward (SF), were applied to eliminate irrelevant information and optimize the efficiency of the machine learning algorithms. Logistic Regression (LR), Nave Bayes (NB), and Random Forest (RF) classifiers, along with ensemble soft voting, were utilized. The experimental results indicated that Recursive Feature Elimination combined with ensemble soft voting achieved the highest accuracy of 97% using the Mendeley dataset and 81% using the Pima Indians Diabetes Dataset.