Development of Machine Learning Models to Predict Hypoglycemia and Hyperglycemia on Days of Hemodialysis in Patients with Diabetes based on Continuous Glucose Monitoring

Background/Objectives: Patients with diabetes undergoing hemodialysis (HD) are at risk of asymptomatic hypo- and hyperglycemia within 24 hours of dialysis. Continuous glucose monitoring (CGM) can improve glycemic control, and machine learning offers a promising approach to detect and predict glycemic excursions based on CGM data. This study aimed to develop machine learning models to predict substantial hypo- and hyperglycemia on dialysis days using CGM data and baseline characteristics. Methods: Using data from 21 patients with diabetes receitarving HD, three classification models (Logistic Regression, XGBoost, and TabPFN) were trained and tested. Predictive features included CGM-derived metrics, HbA1c levels, and insulin use. A binary classification approach was used to predict level 2 hyperglycemia and level 1 hypoglycemia based on international consensus targets; CGM derived Time Above Range (TAR) [≥]10% and Time Below Range (TBR) [≥]1%. Results: A total of 555 dialysis days were included in the analysis. The Logistic Regression model achieved the best performance for predicting hyperglycemia (F1 score: 0.85 [CI95,0.75-0.91]; ROC-AUC: 0.87 [CI95,0.78-0.93]). For hypoglycemia, TabPFN performed best (F1 score: 0.48 [CI95,0.26-0.69]; ROC-AUC: 0.88 [CI95,0.77-0.94]). Conclusion: Prediction of substantial hypo- and hyperglycemia in patients with diabetes undergoing HD appears feasible using machine learning models. Additional studies are needed to confirm clinical utility and generalizability.