A machine learning model to predict intradialytic hypotension in pediatric continuous kidney replacement therapy

Abstract

Background: Intradialytic hypotension (IDH) is associated with mortality in adults undergoing intermittent hemodialysis, but this relationship is unclear in critically ill children receiving continuous kidney replacement therapy (CKRT). We aim to evaluate the relationship between IDH and hospital mortality and if pressure data from dialysis machines could predict IDH.

Methods: We conducted a retrospective cohort study in a tertiary pediatric intensive care unit and NICU from December 2019 to July 2022, including 23 patients across 38 admissions (median age 10 years).

Results: IDH proportion was significantly associated with mortality (risk ratio [RR]: 4.40, 95% confidence interval [CI]: 1.22-15.90, p = 0.02). Random Forest models using Entropy or Gini criteria demonstrated high sensitivity. The CatBoost model achieved the highest average F1-score and area under the receiver operating characteristic (ROC) curve (AUC) (88.18% and 86.6% with and without dialysis settings, respectively). Local Interpretable Model-agnostic Explanations (LIME) indicated that dialysis machine-derived time-series pressure parameters were critical predictive features for IDH, whereas blood pressure-related variables were not among the top predictors.

Conclusions: Dialysis machine-derived pressure parameters may serve as effective predictive markers for IDH, which is associated with increased mortality. These findings support the potential of integrating pressure data in the early detection and management of IDH in pediatric CKRT patients.

Keywords: Continuous kidney replacement therapy; Intradialytic hypotension; Machine learning; Mortality.