Prediction of early-stage diabetes risk with machine learning and statistical techniques

Machine learning and statistical methods contribute to the accurate and meaningful analysis of big data in healthcare, economics, and many other sectors, thereby relieving the workload. In this thesis, machine learning-based early diagnosis models for diabetes, which exhibit high accuracy prediction rates, were developed with the aim of contributing to the existing literature. It is expected that these models will effectively assist in clinical decision-making processes in early diagnoses within the healthcare sector. In the thesis, overcoming the challenges associated with obtaining health data in larger quantities has been facilitated through synthetic data. Additionally, machine learning-based approaches have been proposed to investigate factors contributing to early diabetes risk and to assist in predicting various conditions using specific clinical findings. Furthermore, statistical analyses of demographic variables in the dataset have been included. The thesis extensively discusses the small size and class imbalance of the dataset used in the study. Strategies for class imbalance such as SMOTE, ROS, RUS, and ADASYN are explained concerning their performance on the early diabetes diagnosis prediction problem. Results indicate that balancing the dataset with SMOTE and ADASYN generally improved performance metrics (accuracy, F1 score, Kappa, precision, Roc AUC) compared to models trained on the original dataset. Particularly high accuracy values were obtained for ANN, SVM, and RF models. The SVM algorithm achieved a 0.99 accuracy value with ADASYN and 0.98 with SMOTE. ANN and RF achieved a 0.99 accuracy value with ADASYN. These results demonstrate that ADASYN and SMOTE methods generally enhance classification performance on imbalanced datasets.