An interpretable patient-level model predicts intradialytic hypoglycemia in maintenance hemodialysis patients with diabetic kidney disease

Abstract

Aims: Intradialytic hypoglycemia is an underrecognized patient-safety problem among patients with diabetic kidney disease receiving maintenance hemodialysis. We developed and internally validated an interpretable electronic health record (EHR)- and dialysis database-based patient-level model intended to support digital risk stratification for 6-month risk of documented intradialytic hypoglycemia using routinely available clinical, biochemical, and dialysis-related variables. Methods: This retrospective single-center prediction-model study included adult patients with diabetic kidney disease who underwent maintenance hemodialysis. The index date was the first eligible hemodialysis session during the study period. Candidate predictors were measured within 30 days before the index date, and the primary outcome was at least one documented intradialytic hypoglycemic event within 6 months, defined as capillary blood glucose <3.9 mmol/L during a hemodialysis session. LASSO logistic regression was used for feature selection, and logistic regression, random forest, XGBoost, LightGBM, and support vector machine models were compared. Results: Among 400 patients, 194 (48.5%) experienced documented intradialytic hypoglycemia. LASSO retained eight predictors: sex, diabetes duration, dialysis vintage, pre-dialysis blood glucose, serum albumin, parathyroid hormone, serum phosphorus, and ultrafiltration volume. In internal validation, logistic regression achieved the best overall performance, with an area under the receiver operating characteristic curve of 0.870 (95% CI, 0.800–0.926), accuracy of 0.783, sensitivity of 0.879, specificity of 0.694, F1-score of 0.797, and Brier score of 0.145. Calibration was acceptable, and exploratory decision curve analysis suggested positive net benefit across a range of threshold probabilities. Conclusion: A parsimonious logistic regression model using routinely available variables showed favorable internal validation performance for patient-level risk stratification of documented intradialytic hypoglycemia. It estimates 6-month patient-level risk and is not intended to predict hypoglycemia during an individual dialysis session. However, because the model was derived from a single-center retrospective cohort and the proposed digital workflow was not implemented or prospectively evaluated, it should be regarded as a preliminary risk-stratification framework rather than a validated clinical decision-support tool. External prospective validation is required before clinical deployment.