The Effect of Combining Datasets in Diabetes Prediction Using Ensemble Learning Techniques

Abstract

Diabetes prediction models often suffer from limited generalizability due to reliance on singlepopulation datasets, which fail to capture the diversity of real-world patient demographics. This limitation reduces their clinical applicability across different ethnic groups and geographic regions. The research aims to improve diabetes prediction accuracy and generalizability by combining multiple datasets and employing ensemble learning techniques, addressing the challenges of imbalanced data and population diversity. The research combines two publicly available datasets (Pima Indians: 768 samples and German Society: 2,000 samples) and utilizes preprocessing procedures conducted on these datasets. By comparing the performance of the individual dataset (Pima Indians and German Society datasets) and the combined dataset, it is clear that the models trained on the combined data show improved performance on all metrics. The Random Forest model outperforms the other ensemble models in the Pima Indians dataset, achieving an accuracy of 0.817. The models with the highest accuracy on the German Society dataset are Gradient Boosting and Random Forest, with respective accuracies of 0.996 and 0.994. Then, in the combined dataset, Gradient Boosting and Random Forest yield the best accuracy of 0.991 and 0.988, respectively. It is noticeable that this improvement reflects the ability of models trained on combined data to better accommodate diversity in the data, allowing them to generalize more effectively when applied to different populations. Future research should explore deep learning techniques and additional diverse datasets to enhance model performance further.