Enhanced Multi-Class Hybrid Machine Learning Techniques Model for Predicting Diabetes Mellitus

Abstract

A widespread chronic condition known as diabetes, poses significant healthcare challenges globally. Machine Learning (ML) algorithms have been shown to aid in enhancing the predictive models of diabetes, yet, it is comparatively difficult for single model approaches to provide maximum probable predictive accuracy. For early intervention, proper management of the condition and prevention of delayed treatment, accurate prediction of a greater number of true positive cases is necessary. This research aims to create a strong and consistent hybrid model for diabetes occurrence prediction with high recall value. Employing an ensemble of various ML algorithms through the utilization of voting technique, this blended model combines the predictive power of Gradient Boosting, Decision Trees, Naïve Bayes, ANN, Random Forest, and Logistic Regression and addresses missing values through the Simpleimputer function using median imputation, outliers through Winzorisation, class imbalance through Balanced Bagging Classifier, and hyperparameter optimization through Grid-search. It applies object-oriented analysis and design methodology. Programming language, Python, is used to accessed the Jupyter Notebook IDE for implementation. Evaluation metrics used to assess the hybrid model are recall, Micro-average, precision, macro-averages, F1-score, AUC and accuracy. The model provides accuracy of 0.77, a recall value of 0.90. Using a voting approach, the study maximizes the predictive accuracy for prediction of diabetic disease using the exclusive strengths of such heterogeneous algorithms. Diabetes outcome predictions are made easier if ML algorithms are used wisely along with a soft voting mechanism.