Development and Validation of DIANA (Diabetes Novel Subgroup Assessment tool): A web-based precision medicine tool to determine type 2 diabetes endotype membership and predict individuals at risk of microvascular disease

Background: Previous research has identified four distinct endotypes of type 2 diabetes in Asian Indians, which include Severe Insulin Deficient Diabetes (SIDD), Combined Insulin Resistant and Deficient Diabetes (CIRDD), Insulin Resistance and Obese Diabetes (IROD), and Mild Age-related Diabetes (MARD). DIANA (Diabetes Novel Subgroup Assessment) is an online precision medicine tool that can predict endotype membership of type 2 diabetes and individual risk for retinopathy and nephropathy. Methodology: The DIANA tool determines subgroup membership using a machine learning model (support vector machine) on T2D subgroups in the Asian Indian population. We used a support vector machine (SVM) model to classify type 2 diabetes patient endotypes, and the model is trained based on k-fold cross-validation. Its performance was compared with an algorithm determined based on conditional pre-determined cut-offs and weights for each clinical feature [age at diagnosis, BMI, waist, HbA1c, Serum Triglycerides, HDL-Cholesterol, (C-peptide fasting, C-peptide stimulated) – optional. This study employed local interpretable model-agnostic explanations (LIME) and SHapley Additive exPlanations (SHAP) to demystify the endotype prediction model. A random forest model was built to assess an individual’s risk for nephropathy and retinopathy based on individual risk algorithms. Findings: The SVM model has relatively high accuracy, specificity, sensitivity, and precision values compared to conditional pre-determined cut-offs 98% vs 63.6%, 99.8% vs 88%, 98.5% vs 65.1%, and 98.7% vs 63.4%. Clinician face value validation of the prediction by the SVM model reported an accuracy, specificity, sensitivity and precision compared to conditional pre-determined cut-offs 97% vs 85%, 95.3% vs 63%, 95.8% vs 73%, and 98.9% vs 66.9%. Additionally, our study demonstrated the impact of features on ML models through LIME and SHAP analyses. The accuracy of the random forest risk prediction model for nephropathy and retinopathy was 89.6% (p