Predicting and Modeling Recovery Dynamics Post-Acute Kidney Injury: A Multicenter Study.

Area of Science:

• Nephrology
• Data Science in Healthcare
• Clinical Informatics

Background:

• Acute kidney injury (AKI) impacts 10-25% of hospitalized patients, leading to significant morbidity and mortality.
• AKI recovery varies, from full reversal to chronic kidney disease progression.
• Predicting short-term AKI states is crucial for optimizing clinical interventions.

Purpose of the Study:

• To develop predictive models for short-term acute kidney injury (AKI) progression and reversal.
• To characterize dynamic AKI recovery and progression patterns using multistate modeling.
• To identify key clinical variables that predict AKI outcomes.

Main Methods:

• Retrospective analysis of 94,531 inpatient encounters from four healthcare systems (2009-2022).
• Development of CatBoost machine learning models to predict 7-day AKI progression and reversal.
• Multistate modeling to estimate transition intensities and covariate effects on AKI state changes.

Main Results:

• Models demonstrated strong predictive performance (AUROC 0.79-0.93) for AKI reversal and progression.
• Serum creatinine, systolic blood pressure (SBP), and albumin were key predictors.
• Low SBP predicted progression; high SBP predicted faster but asymmetric recovery. Nearly half of AKI-1 patients recovered within 1 day.

Conclusions:

• A two-stage framework effectively predicts and characterizes early AKI recovery and progression.
• Routinely measured variables like serum creatinine and blood pressure trends track AKI dynamics.
• Findings require prospective validation but support refining risk stratification and targeted AKI interventions.