Type 2 Diabetes Prediction Using Recall-Optimized XGBoost

Early detection of Type 2 Diabetes (T2D) is essential to avoid long-term consequences and minimize the burden on health care. This paper documents the development of a machine learning-based predictive model using the Pima Indians Diabetes Dataset and the XGBoost algorithm. Data preprocessing, class imbalance handling using SMOTE, and hyperparameter optimization through randomized search were carried out to improve model performance. The final model achieved 74.6% accuracy, a recall of 83%, and an ROC-AUC of 0.829, indicating high sensitivity and strong class discrimination. This study emphasizes the importance of optimizing recall in clinical diagnosis settings, where minimizing false negatives is more critical than maximizing overall accuracy. Compared to existing research that primarily prioritizes accuracy, this study addresses the clinical need for sensitivity in screening tools. The results demonstrate the potential of recall-optimized machine learning models to enhance early detection strategies in primary care, especially within resource-limited environments.