Applying Supervised Machine Learning to Predict Diabetes from E-Health Records

Abstract

Diabetes mellitus remains a major global health issue, particularly in the United States, where complications from persistent hyperglycemia contribute to severe health outcomes. Although prior research has improved diabetes prediction models, there is still a need for more accurate and efficient machine learning approaches to support early diagnosis and timely intervention. This study explores early diabetes detection using machine learning techniques applied to electronic health records. Four supervised algorithms—Random Forest, Gradient Boosting, Support Vector Classifier (SVC), and K-Nearest Neighbors (KNN) were compared to identifying the most relevant predictive features and determining which model performs best across training, testing, and validation phases. Among the evaluated models, Random Forest and Gradient Boosting demonstrated the strongest and most consistent results, achieving training accuracy of approximately 82.5% and test accuracy around 82.2%. Although the KNN model showed high sensitivity (98.54%) and specificity (93.63%) in a separate assessment, its test accuracy in this study was slightly lower at 77.9%. The SVC model reached a test accuracy of 76.7%. Validation outcomes further confirmed the stability and generalizability of the Random Forest and Gradient Boosting models when applied to unseen data. Overall, the results indicate that ensemble methods deliver more accurate and reliable predictions for diabetes diagnosis compared to KNN and SVC, making them more suitable for clinical settings. While KNN performed well under certain conditions, ensemble techniques provided better generalization. Future work should expand the feature set and apply thorough hyperparameter optimization to improve predictive performance and enable more effective early detection and management of diabetes.