Early Acute Kidney Injury Prediction in Two Multicenter Datasets: Designing for Implementation

Monday, February 24, 2025

1:30pm - 2:30pm Eastern Time

Location: Connections Central Theater 1

Adam Dziorny, MD, PhD

Golisano Children's Hospital University of Rochester

Disclosure information not submitted.
SD

Stephen Drury, MS

Data Analyst
University of Rochester

Disclosure information not submitted.
AC

Alex Clark, MS

Data Analyst
University of Rochester

Disclosure information not submitted.
MZ

Martin Zand, MD, PhD

Professor of Medicine
University of Rochester

Disclosure information not submitted.
Tell Bennett, MD, MS

Professor of Biomedical Informatics and Pediatrics
University of Colorado School of Medicine

Disclosure information not submitted.
RF

Reid Farris, MD, MS

Associate Professor of Pediatrics
Seattle Children's Hospital

Disclosure information not submitted.
Akira Nishisaki, MD, MSCE (he/him/his)

Division of Critical Care Medicine, Department of Anesthesia and Critical Care
The Children's Hospital of Philadelphia

Disclosure information not submitted.
DT

Daniel Tawfik, MD, MS

Assistant Professor of Pediatrics
Stanford University School of Medicine

Disclosure information not submitted.
TC

Tim Cornell, MD

Professor of Pediatrics
Stanford University - Lucille Packard Children's Hosptal Stanford

Disclosure information not submitted.
Sareen Shah, MD

Pediatric Intensivist
Cedars-Sinai Medical Center

Disclosure information not submitted.
Randall C. Wetzel, BS, MB, MBA

Professor of Anesthesiology
Children's Hospital of Los Angeles

Disclosure information not submitted.
Lazaro N. Sanchez-Pinto, MD, MBI

Ann & Robert H. Lurie Children's Hospital of Chicago

Disclosure(s): No relevant financial relationship(s) to disclose.

Disclosure(s):

Lazaro Sanchez-Pinto, MD, MBI: No relevant financial relationship(s) to disclose.

Introduction: Machine learning (ML) can predict adverse events such as acute kidney injury (AKI) in critically ill children. Most models for pediatric AKI prediction are developed on single-center data which limits generalizability. We hypothesize that adding a priori defined features will improve performance compared to a recalibrated existing single-center-derived AKI prediction model (Sanchez-Pinto et al, PCCM, 2016) when tested in two large multicenter pediatric critical care datasets.

Methods: Retrospective cohort study within two datasets: (1) a 5-center dataset linking Virtual Pediatric Systems with PEDSnet (VPS-PN) using probabilistic methods, and (2) the 8-center PICU Data Collaborative (PDC) dataset. We predict AKI within 72 hours of ICU admission, defined by KDIGO serum creatinine criteria. We applied standard ML techniques [e.g., data splitting, imputation, and multiple feature selection techniques] to train, tune, and test new models. We evaluated prediction score effectiveness with multiple standard measures (area under ROC, AUROC; area under precision-recall curve, AUPRC; sensitivity, positive predictive value, PPV, and number needed to alert [NNA] at two specificity thresholds).

Results: Our two datasets included a combined 187,947 ICU encounters, of which 4,797 (2.6%) had AKI within 72 hours of ICU admission. We first recalibrated the existing single-center model and found lower AUROC (0.76 and 0.60) and lower PPV at the 90th percent cutoff (0.08 and 0.10) in both datasets compared to the published single-center results. We trained and tested a new prediction model, selecting a priori defined features and model types based on AUROC and AUPRC targets. Logistic regression model performance on holdout datasets improved compared to recalibrated single center model [AUROC (0.80 and 0.79), sensitivity (0.47 and 0.61), PPV (0.10 and 0.13), and NNA (10 and 8) at the 90th percent cutoff].

Conclusions: We developed new optimal AKI prediction models within two of the largest pediatric critical care datasets and demonstrated improved performance over a recalibrated single-center model. We specifically included only features that are available at the time of prediction. We are implementing this model silently using EHR web interfaces and prospectively monitoring performance.