Diabetes Prediction using Ensemble Learning Techniques

Worldwide, diabetes mellitus is a major hazard to public health, and early detection is crucial for efficient care and issue prevention. In this work, ensemble learning techniques have been proposed for the Behavioural Risk Factor Surveillance System (BRFSS) Health Indicator dataset to propose a complete method for diabetes prediction. Our dataset comprises 253,680 survey responses covering 21 features: medical history, health behaviours, and demographics. The target variable is a binary indication of diabetes’s existence. Employing a method called SMOTE-ENN, which stands for Synthetic Minority Oversampling Technique combined with Edited Nearest Neighbours, can help address issues related to class imbalance. The dataset’s features are revealed through exploratory data analysis (EDA), which directs the following feature engineering process. The authors proposed XGBoost in feature selection to find the most pertinent features to the given task and ensemble of machine learning classifiers (XGBoost, Random Forest, CatBoost, LightGBM, and KNN). Applying the proposed method to the Behavioural Risk Factor Surveillance System (BRFSS) dataset yields an impressive 96.40% accuracy, indicating its efficacy. This high degree of accuracy demonstrates that the approach can predict diabetes. The findings suggest that the method can be helpful for healthcare professionals; feature significance analysis also identifies significant predictors that go into diabetes risk assessment.