Derivation and external validation of machine learning prediction for severe acute kidney injury in patients with sepsis-induced coagulopathy: a study based on the MIMIC database and a Chinese cohort.

Abstract

Sepsis-induced coagulopathy (SIC) is a critical complication of sepsis that profoundly affects the severity of organ dysfunction and clinical prognosis. The current study aimed to construct and compare multiple individualized machine learning (ML) prediction models for predicting severe acute kidney injury (AKI) in patients with SIC, with the goal of identifying the most robust model to improve the accuracy of clinical decision-making. The retrospective analysis was conducted based on the MIMIC‑IV database and a Chinese cohort. Eligible patients with sepsis and SIC were included. Both LASSO regression and the Boruta algorithm were used to select predictive variables. Eight ML models were constructed to predict the individual probability for severe AKI development. The receiver operating characteristic (ROC) curve, area under the precision-recall curve (PR-AUC) value, calibration indicators, and net benefit were used to assess the predictive model performance. Subgroup analyses were performed to evaluate the stability of model performance. Model predictions and feature importance were estimated using SHapley additive explanations (SHAP). Among the 14,155 patients with sepsis, 5,207 developed SIC and were randomly divided into training and internal validation cohorts. The severe AKI incidence rate was 24.7% (1288/5207). The Chinese external validation cohort enrolled 525 patients with SIC, with a severe AKI incidence rate of 28.1% (148/525). Six key predictive variables were identified to construct the ML models. The XGBoost prediction model demonstrated the most robust performance and best discrimination, achieving an area under the ROC curve of 0.875 (0.862-0.888) in the training cohort, 0.831 (0.807-0.854) in the internal cohort, and 0.843 (0.806-0.879) in the external validation cohort. Serum creatinine on the day of SIC diagnosis was identified as the most influential feature in the SHAP analysis. Subgroup analyses demonstrated that the model maintained predictive efficacy when stratified by age, sex, chronic kidney disease, and liver disease. The XGBoost machine learning model performed well in predicting the risk of severe AKI in patients with SIC, exhibiting satisfactory generalizability and interpretability. This prediction model provides a valuable and evidence-based tool for early identification and facilitates personalized targeted management in critical care settings.