Detecting diabetes and identifying it’s physiological indicators are important for advancing medical understanding and improving quality of life for individuals. We explored interpretable machine learning (ML) methods to identify the vital physiological indicators from available three datasets. It was found that our ensemble model outperformed individual models on the two datasets, achieving accuracies of 99.5% and 98.2% respectively. On another dataset, a highest accuracy of 99.3% was achieved using a ResNet-50 model. A SHapley Additive exPlanations (SHAP) method was employed and found that some features, such as glucose, BMI, and age, are useful than others when computing the performances. All the factors that can assist in diagnosing of diabetes, were ranked in this study, and areas contributing to diabetes-induced retinopathy were labeled. These findings underscored the potential of interpretable ML to develop effective diagnostic tools, potentially improving patient outcomes. Overall, we found that our approach achieved highest performance compared to others. The performance highlights the potential of our approach to enhance diabetes diagnosis & management.