Development and external validation of a machine learning model for early identification of hypotension associated with renal replacement therapy in ICU patients

Abstract

Renal replacement therapy (RRT) is essential for critically ill patients in the intensive care unit (ICU), yet hypotension remains its most common complication, increasing the risk of organ hypoperfusion and poor outcomes. No reliable tool currently exists to predict hypotension risk before RRT initiation, particularly one designed for nursing integration. This study aimed t o develop and validate a nurse-friendly machine learning model using readily accessible pre-RRT variables to predict the risk of RRT-associated hypotension in ICU patients. Data from eligible patients in the MIMIC-IV database were used for model development and internal testing, and an independent Chinese ICU cohort was used for external validation. Candidate predictors were identified through an evidence-informed framework combining meta-analysis factors with prespecified patient and treatment variables. Initial predictors were screened according to clinical relevance and routine availability to ICU nurses before RRT initiation. Final feature selection was performed using least absolute shrinkage and selection operator (LASSO) regression in the training set, followed by the development and validation of multiple machine learning models. SHapley Additive exPlanations (SHAP) were used to enhance model interpretability, and the final model was implemented as a web-based prototype. A total of 1,342 patients from the MIMIC-IV database and 133 patients from the external set were included in the analysis. The gradient boosting machine (GBM) demonstrated the most consistent overall performance across all datasets, with an area under the curve (AUC) of 0.801 in the external set. Key contributors to model output included RRT modality, vasopressor use, mean and systolic blood pressures, age, lactate level, and the interval from ICU admission to RRT initiation. The final model was deployed using the Streamlit framework to demonstrate individualized risk visualization. The GBM-based model demonstrated acceptable predictive performance, nursing applicability, and clinical utility. Designed with a nursing-friendly approach, the model enables early risk stratification and supports proactive hemodynamic management before RRT initiation. Not applicable.