Machine Learning-Based Subtype Classification of Sepsis-Associated Acute Kidney Injury and Differential Responses to Renal Replacement Therapy

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Background Sepsis-associated acute kidney injury (SA-AKI) is highly heterogeneous with controversial optimal renal replacement therapy (RRT) strategies. Machine learning enables subphenotype identification, and exploring RRT response differences across SA-AKI subtypes is critical for precision care. Methods This retrospective cohort study analyzed adult SA-AKI patients from the Medical Information Mart for Intensive Care IV database. K-means clustering classified SA-AKI subtypes; nine machine learning models were built to predict 28-day mortality, with model performance validated and interpreted via Shapley Additive Explanations. Multivariable logistic regression with interaction terms assessed RRT's heterogeneous treatment effects (HTE) across subtypes. Sensitivity analysis was used to assess the reliability of the research results. Results 21359 patients were stratified into 3 AKI subtypes (C1: moderate severity; C2, severe hyperglycemia; C3, mild conditions and high inflammatory responses). C2 had the highest 28-day mortality (30.8%) and RRT utilization (19.2%). Random Forest had a greater performance in training set [area under the curve (AUC) = 0.961], while Light Gradient Boosting Machine (Lightgbm) was better in test set (AUC = 0.814), with Sequential Organ Failure Assessment (SOFA) score as the top predictor. RRT had a negative correlation with 28-day mortality [odds ratio (OR)= 0.810, P = 0.006], with a significant difference only in C1. HTE test for RRT showed no statistical significance ( P = 0.141). Sensitivity analysis results supported the reliability of the research findings. Conclusion SA-AKI has three distinct subtypes with divergent clinical characteristics. Lightgbm effectively predicts SA-AKI prognosis, and RRT may benefit C1 subtype patients. Although there is not enough evidence to prove the HTE of RRT among subgroups, this study provides a phenotypic framework for advancing precision kidney support in SA-AKI.