Artificial Intelligence Models for Predicting Acute Kidney Injury in Adult Critical Care: A Systematic Review

Abstract

Introduction. Acute kidney injury (AKI) is frequent in adult critical 
care and is associated with increased morbidity, mortality, and 
long-term kidney consequences. Artificial intelligence (AI) and 
machine-learning (ML) prediction models may enable earlier risk 
identification, but clinical adoption depends on robust reporting 
beyond discrimination, including calibration and clinical utility.
Methods. We performed a PubMed-based systematic review 
(English; last 5 years) of AI/ML models predicting AKI in adult 
critical care/ICU populations. Records were screened and full 
texts were assessed for eligibility. For synthesis within the journal 
page limit, we restricted the final evidence set to studies reporting 
a complete usability-oriented outcome set: AKI definition, model 
type(s), prediction horizon, validation approach, AUROC/AUC, 
calibration, and decision curve analysis (DCA)/net benefit.
Results. Of 357 records screened, 230 full texts were assessed, 
and 31 studies met complete-reporting criteria and were included. 
The most common model families included logistic regression 
baselines, random forest, gradient boosting/XGBoost, and deep 
learning. Reported AUROC/AUC values ranged from 0.64 to 
1.00 (median 0.90), with the best-performing models typically 
reporting AUROC ≥ 0.95. External or temporal validation was 
reported in 74% (23/31). By design, 100% of included studies 
reported calibration (e.g., calibration plot and/or Brier score 
and/or Hosmer–Lemeshow) and 100% reported clinical utility 
using DCA/net benefit.
Conclusions. Among usability-focused studies, AI/ML models 
show generally strong discrimination for AKI prediction in 
adult critical care, but reporting and validation practices remain 
heterogeneous. Standardized AKI definitions, transparent 
validation, calibration reporting, and decision-analytic evaluation 
are essential to support safe implementation.