A Comparative Performance Analysis of Hybrid and Classical Machine Learning Method in Predicting Diabetes

Diabetes mellitus is one of medical science’s most important research topics because of the disease’s severe consequences. High blood glucose levels characterize it. Early detection of diabetes is made possible by machine learning techniques with their intelligent capabilities to accurately predict diabetes and prevent its complications. Therefore, this study aims to find a machine learning approach that can more accurately predict diabetes. This study compares the performance of various classical machine learning models with the hybrid machine learning approach. The hybrid model includes the homogenous model, which comprises Random Forest, AdaBoost, XGBoost, Extra Trees, Gradient Booster, and the heterogeneous model that uses stacking ensemble methods. The stacking ensemble or stacked generalization approach is a meta-classifier in which multiple learners collaborate for prediction. The performance of the homogeneous hybrid models, Stacked Generalization and the classic machine learning methods such as Naive Bayes and Multilayer Perceptron, k-Nearest Neighbour, and support vector machine are compared. The experimental analysis using Pima Indians and the early-stage diabetes dataset demonstrates that the hybrid models achieve higher accuracy in diagnosing diabetes than the classical models. In the comparison of all the hybrid models, the heterogeneous model using the Stacked Generalization approach outperformed other models by achieving 83.9% and 98.5%.