Cohort study uncovers three clinical subtypes and differential dialysis responses in sepsis-associated acute kidney injury, indicating potential for personalized therapy.
Key Points
To identify clinical subtypes of sepsis-associated acute kidney injury using machine learning and evaluate whether response to renal replacement therapy varies across these subphenotypes.
Analyzed retrospective cohort data from 21,359 adult patients with sepsis-associated acute kidney injury in the MIMIC-IV database.
Applied K-means clustering to classify disease subtypes and trained nine machine learning models, interpreted via Shapley Additive Explanations, to predict 28-day mortality.
Assessed heterogeneous treatment effects of renal replacement therapy using multivariable logistic regression with interaction terms alongside sensitivity analyses.
Stratified patients into three clinical subtypes: C1 (moderate severity), C2 (severe hyperglycemia, highest 28-day mortality at 30.8% and RRT use at 19.2%), and C3 (mild conditions with high inflammation).
Light Gradient Boosting Machine achieved the strongest test set discrimination (AUC = 0.814), with the Sequential Organ Failure Assessment score identified as the leading mortality predictor.
Renal replacement therapy was associated with lower 28-day mortality overall (OR = 0.810, P = 0.006) and specifically within C1, though the overall interaction test for heterogeneous treatment effect was not statistically significant (P = 0.141).