Optimizing diabetes prediction accuracy: A comprehensive approach with advanced preprocessing and diverse machine learning classifiers

Abstract

Diabetes is a prevalent chronic health disease affecting millions of people over the world. Early detection and effective management contribute in preventing its complications. This study investigates the use of machine learning algorithms to predict the risk of acquiring diabetes based on multiple clinical parameters (Indian PIMA Dataset). In the study, advanced data preprocessing techniques like missing data handling, outlier rejection, robust scaling and SOMTE oversampling are consolidated with eight traditional machine learning. Effectiveness of those eight models such as Gradient Boosting (GB), XGBoost (EGB), Random Forest (RF), Light GBM (LGBM), Cat Boost (CB), Ada Boost (AB), Decision Tree (DT) and K-Nearest Neighbor (KNN) are analyzed and compared. The prediction outcomes of each model are contrasted using accuracy, roc score and F1 score. After evaluating the model's performances, GB has provided the highest accuracy of 90.25%, ROC score of 96.38% and F1 score of 86.48%.