Comorbidity prediction model for Type 2 Diabetes Patients - Survival-based framework

‍

‍

Abstract:

Assessing the risk of complications among Type-2 Diabetes Mellitus (T2DM) patients is a crucial step for providing effective care for T2DM patients. Clinical Decision Support Systems (CDSS) can help clinicians take advantage of available knowledge contained within a hospital's EHR system to estimate individual risk. In this study, we develop and implement a Machine Learning (ML) based survival model suite for three major comorbidities of T2DM: ArtheoSclerotic CardioVascular Disease (ASCVD), End Stage Renal Disease (ESRD), and Sight Threatening Diabetic Retinopathy (STDR). Our modeluse survival XGBoost architecture, and can handle right censored data. Model performance was assessed via the c-statistic showing a mean patient-level prediction performance of 0.72, 0.92, and 0.76 respectively (3 fold cross validation). Additionally, we apply SHapley Additive exPlanations (SHAP) analysis to explain the models’ risk predictions.