Machine Learning for Early Non-invasive Diabetes Detection Using Electronic Health Records

Early detection of Type 2 Diabetes Mellitus (T2DM) is critical to preventing long-term complications such as cardiovascular disease, nephropathy, and retinopathy. However, conventional diagnostic approaches are often invasive, costly, and unsuitable for population-scale screening. This study proposes a non-invasive, machine learning-based framework for early T2DM detection using electronic health records (EHRs) from a publicly available Kaggle dataset. Key non-invasive features including demographics, vital signs, medication history, and temporal health trends were extracted and used to train six classifiers: random forest (RF), support vector machine (SVM), naïve bayes (NB), alternating decision tree (ADT), random tree (RT), and k-nearest neighbors (KNN). Class imbalance was addressed using the synthetic minority over-sampling technique (SMOTE) at 0%, 150%, and 300% levels. Experimental results show that RF achieved the highest AUC (88.45%) at 150% SMOTE, while SVM demonstrated the best sensitivity gains when temporal features and feature selection were applied. The proposed framework demonstrates the potential of interpretable, EHR-based ML models for scalable, cost-effective diabetes screening and offers a reproducible benchmark for future applications in real-world clinical data.