Predicting Glucose Levels in Diabetic Kidney Transplant Recipients Using Multivariate Temporal Modeling Across Clinically Defined Subcohorts

Metabolic complications following kidney transplantation, particularly dysglycemia, pose a significant challenge, especially among diabetic kidney transplant recipients. This condition is largely driven by the effects of immunosuppressive medications, particularly corticosteroids such as prednisone, which disrupt glucose regulation and increase the risk of post-transplant complications. Effective monitoring and timely medication adjustments are crucial for optimal glycemic control in these individuals. In this study, we developed a predictive model for glucose levels by integrating continuous glucose monitoring data with clinical, demographic and medication information. Through clustering analysis, we identified five distinct patient clusters, and then applied subcohort-specific vector autoregressive models to capture the interaction between bolus insulin and glucose. Our approach showed a strong predictive performance for next-day glucose levels, with a normalized root mean square error (NRMSE) below 8.8% ($RMSE \lt 30 \mathrm{mg} / \mathrm{dL}$) and an average correlation greater than 0.80 between predicted and actual values. In addition, most cluster-specific models exhibited an improvement over the generic model, which was trained on the entire dataset. Importantly, our approach offers valuable insights for personalized insulin dosing, as we detected differing associations between glucose levels and insulin on the identified subcohorts. Overall, we consider our approach valuable for digital health solutions aimed at post-transplant diabetes management.