Diabetes and stroke are two diseases that are closely interconnected with chronic conditions affected by shared metabolic and vascular risk factors.. This study presents a comparative analysis of five machine learning models, which include CatBoost, LightGBM, XGBoost, as well as Random Forest and Logistic Regression, to predict diabetes and stroke based on structured electronic health record data. Accuracy, Precision, Recall, F1-score, and ROC-AUC were used to evaluate model performance where balanced weighting methods corrected performance imbalance between classes. Experimentally, it was found that the ensemble gradient boosting models performed better than the traditional methods with the highest performance of CatBoost on diabetes prediction (Accuracy: 98.42, F1-score: 0.9842) and strong results on stroke prediction (Accuracy: 94.10, F1-score: 0.9408). In addition to performance comparison, the study compares feature relationships by applying Pearson correlation analysis and SHAP-based interpretability. Pearson analysis has also revealed that HbA1c was the best predictor of diabetes (r=0.85) but age and hypertension were more correlated with stroke. SHAP ranking also emphasized the predominant role of glucose-related and vascular factors in each of the prediction tasks and combined statistical correlation and explainable AI methods to gain a better insight into significant predictive characteristics of the underlying determinants of diabetes and stroke risk.