Assessing hypoglycemia risk during hemodialysis using an explainable machine learning approach based on continuous glucose monitoring metrics

Continuous glucose monitoring (CGM) can identify hypoglycemia in hemodialysis (HD) patients, who are at risk for this event. On the other hand, machine learning has remarkable value in CGM-based studies, but none of previous studies addressed the problem of hypoglycemia prediction in HD patients. Therefore, we conducted this study to setup different machine learning models based on CGM data (specifically, CGM metrics) to assess the risk of mild and severe hypoglycemia during HD sessions. We studied a cohort of twenty patients (11 with and 9 without diabetes) undergoing chronic HD. All patients underwent CGM for up to 2 weeks. We identified 92 HD sessions and related pre-HD sessions of 8 h length. HD sessions were used to identify mild (<70 mg/dL) and severe (<54 mg/dL) hypoglycemia, whereas pre-HD sessions were used to compute 48 CGM metrics. We then performed feature selection to identify the most relevant metrics for hypoglycemia prediction. The metrics performance was assessed with binary decision tree, k-nearest neighbors, penalized logistic regression, Na & iuml;ve Bayes, random forest ensemble algorithm. We found that mild hypoglycemia was best predicted by six metrics (M-value(100), TIR70-180, ADRR, MAGE-, MAG(30), CONGA(1h)), whereas severe hypoglycemia by three metrics (TIR70-180, ADRR, CONGA(1h)). The best overall performance was achieved by the tree, showing area under receiver operating characteristic curve (AUC) equal to 68.2 % for prediction of mild hyperglycemia, and AUC equal to 81.2 % for severe hyperglycemia. Notably, individual hypoglycemia risk assessment has potential to guide personalized HD-related clinical decisions to minimize such risk.