Improved Diabetes Prediction: Advanced Hybrid Sampling Method of Kmeans-SVMSMOTE-ENN and Ensemble Learning Model

Over the decades, diabetes has always been an important health condition in the world that has grown rapidly with a tremendous impact. In recent years, machine learning has become a valuable tool for early disease prediction especially diabetes; however, there remain limitations in the area from previous literature, with the challenges of imbalanced data, a single classification model, and an inadequate dataset. In this regard, this paper presents a new advanced hybrid sampling method and a proposed framework with an ensemble learning model to address these research gaps. The new sampling method provides a well-rounded structure by starting with SVM-SMOTE (Support Vector Machine Synthetic Minority Over-Sampling Technique) according to K-means clustering, then cleaning with ENN (Edited Nearest Neighbors) under-sampling for better quality and relevant balance data. Based on the comparison with the other seven resampling methods, this advanced hybrid sampling method has shown the best performance. Moreover, there is also a development of ensemble model investigation in max voting with three different types of models, including k-nearest neighbor (KNN), decision tree (DT), and categorical boosting (CatB). To indicate a comprehensive evaluation, experimental studies are conducted with six distinct circumstances of diabetes datasets by using the metrics of accuracy, precision, recall, and F1 score. The experiment results showed that the novel research framework with Kmeans-SVMSMOTE-ENN and the ensemble model has achieved a great outcome that outperforms other studies, which demonstrates the validation of the method and the strong efficiency of the proposed framework in the case study of diabetes early detection.