Abstract: SA-PO0685
Artificial Intelligence for Predicting Pediatric AKI: Systematic Review and Meta-Analysis
Session Information
- Pediatric Nephrology: Transplantation, Hypertension, AKI, Genetics, and Developmental Diseases
November 08, 2025 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Pediatric Nephrology
- 1900 Pediatric Nephrology
Authors
- Shirode, Parth S., Akron Children's Hospital, Akron, Ohio, United States
- Raina, Rupesh, Akron Children's Hospital, Akron, Ohio, United States
- Cheungpasitporn, Wisit, Mayo Clinic Minnesota, Rochester, Minnesota, United States
Background
Acute kidney injury (AKI) in hospitalized children is a major complication associated with significant morbidity and mortality. Current diagnostic criteria, relying on serum creatinine and urine output, are delayed and often imprecise. The integration of artificial intelligence (AI)/machine learning (ML) models may enable early detection and risk stratification. This systematic review evaluates the performance of AI/ML models for predicting pediatric AKI across clinical settings.
Methods
We searched PubMed, Embase, and Web of Science for studies applying AI/ML models to predict AKI in pediatric populations. Studies reporting performance metrics such as AUC, sensitivity, specificity, PPV, NPV, accuracy, and F1 score were included. Data extraction and risk of bias assessments were performed independently by two authors.
Results
Among 470 records identified, 11 studies met the inclusion criteria, with 14 AI/ML models used. The overall sample size included 33,949 children with an AKI proportion of 12.5%. Meta-analyses of AUC were conducted on neural network, gradient boosting, and logistic regression. Gradient boosting had the highest pooled AUC of 0.873 (95% CI: 0.836-0.909). Random forest demonstrated the highest median sensitivity (0.821), specificity (0.942), PPV (0.860), NPV (0.935), and accuracy (0.821).
Conclusion
Gradient boosting, random forest, and logistic regression demonstrated better predictive performance. However, small sample size, heterogeneity, lack of testing/validation cohorts, insufficient data, and inconsistent patient populations and AKI diagnostic criteria restricts generalizability. Standardized data collection, inclusion of age-specific biomarkers, and comprehensive feature engineering could enhance the effectiveness of these models.