Enhanced diabetes prediction using SMOTE-Tomek and F1-optimized machine learning models

Abstract

This study presents a systematic and effective strategy for the early prediction of diabetes, utilizing a hybrid machine learning framework designed to address prevalent issues found in conventional diagnostic systems. A significant challenge in medical datasets, particularly in the context of diabetes, is the issue of class imbalance, characterized by a disproportionate representation of non-diabetic cases compared to diabetic cases. The proposed framework addresses this issue by implementing the synthetic minority oversampling technique (SMOTE)-Tomek resampling in conjunction with a model selection strategy that emphasizes the F1-score for the diabetic class. This approach ensures a heightened focus on accurately identifying at-risk individuals, rather than solely optimizing for overall accuracy. A variety of supervised learning models were assessed, comprising logistic regression (LR), decision trees (DT), support vector machines (SVM), random forest (RF), k-nearest neighbors (KNN), extreme gradient boosting (XGBoost), and artificial neural networks (ANN). A 10-fold cross-validation technique was utilized to ensure a robust evaluation of model performance. The evaluated models, ANN and XGBoost, demonstrate notable performance, with F1-scores of 0.87 and 0.84, respectively. The ANN model achieves the highest accuracy, recorded at 94.3%. SMOTE-Tomek has enhanced recall and decreased false negatives, thereby increasing the model's sensitivity to diabetic cases.