Work

Predicting and Modeling Recovery Dynamics Post–Acute Kidney Injury

Year: 2025

Type: article

Abstract: Key Points Machine learning models predicted short-term AKI reversal and progression with area under the receiver operating characteristic curves of 0.79–0.93 across four health systems. Top predictors common to both outcomes included serum creatinine, systolic BP, and albumin, with consistent patterns across diverse sites. Early AKI recovery is often transient, with frequent relapse, highlighting the need for continued monitoring after initial improvement. Background AKI affects 10%–25% of hospitalized patients and is associated with significant morbidity, mortality, and long-term adverse outcomes. Recovery patterns vary greatly, from full reversal to progression toward CKD. Predicting short-term AKI states and understanding their dynamic trajectories could help optimize clinical interventions. Methods This retrospective study used data from four health care systems in the Greater Plains Collaborative, spanning 2009–2022. A total of 94,531 inpatient encounters (54% male, 80% White patients) from 75,312 adults with admission eGFR ≥15 ml/min per 1.73 m 2 and without prior dialysis or kidney failure were included. We developed CatBoost models, a gradient-boosting decision-tree algorithm, to predict 7-day AKI progression and reversal and then performed multistate modeling to estimate transition intensities and covariate effects on AKI state changes. Results The models achieved strong performance across outcomes and subgroups: area under the receiver operating characteristic curve 0.79–0.89 and accuracy 0.7–0.79 for reversal; area under the receiver operating characteristic curve 0.91–0.93 and accuracy 0.83–0.87 for progression. Serum creatinine, systolic BP (SBP), and albumin were key predictors across both outcomes. Low SBP (≤120 mm Hg) predicted increased likelihood of progression, whereas high SBP (≥140 mm Hg) predicted asymmetric recovery with reduced chances of subsequent reversal. Multistate analysis highlighted the dynamic and transient nature of AKI recovery, with nearly half of AKI-1 patients (instantaneous hazard rate=0.46) transitioning to “no AKI” peaking on the first day postonset. Conclusions This study presents a two-stage analytical framework linking short-term prediction with multistate modeling to characterize early recovery and progression patterns after hospital-acquired AKI. The approach highlights a small set of routinely measured variables, such as serum creatinine trajectories and BP trends, that consistently track these dynamics across health systems. While the results are hypothesis-generating and require external, prospective validation, they lay groundwork for future studies on refining risk stratification and exploring targeted interventions in AKI care.

Source: Kidney360

DOI: 10.34067/kid.0000001052