Abstract: FR-PO035
Predicting Early AKI in Two Large Multicenter Pediatric Critical Care Datasets
Session Information
- AKI: Epidemiology, Risk Factors, and Prevention - 2
October 25, 2024 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Acute Kidney Injury
- 101 AKI: Epidemiology, Risk Factors, and Prevention
Authors
- Dziorny, Adam, Golisano Children's Hospital, Rochester, New York, United States
- Drury, Stephen C., University of Rochester Medical Center, Rochester, New York, United States
- Clark, Alex, University of Rochester Medical Center, Rochester, New York, United States
- Zand, Martin S., University of Rochester Medical Center, Rochester, New York, United States
- Sanchez-Pinto, L. Nelson N., Northwestern University Feinberg School of Medicine, Chicago, Illinois, United States
Background
Machine learning (ML) can predict adverse events such as acute kidney injury (AKI) in critically ill children, allowing for proactive care strategies. Most models for pediatric AKI prediction are developed on single-center data which limits generalizability. In this study we externally validate an existing single-center-derived AKI prediction model, recalibrate, and add features on two of the largest multicenter pediatric critical care datasets available.
Methods
Retrospective cohort study within two datasets: (1) a 5-center dataset linking Virtual Pediatric Systems with the PEDSnet (VPS-PN) using probabilistic methods, and (2) the 8-center PICU Data Collaborative (PDC) dataset. We predict AKI within 72 hours of admission, defined by KDIGO serum creatinine criteria. We applied standard ML techniques, including data splitting, imputation, feature selection, and model training, to develop and cross-train a new model. We evaluated prediction score effectiveness with multiple standard measures (area under ROC, AUROC; area under precision-recall curve, AUPRC; and PPV [precision] at two sensitivity thresholds).
Results
Our two datasets included a combined 190,054 ICU encounters, of which 4,797 (2.5%) had AKI within 72 hours of ICU admission. We first applied the existing single-center model and found lower AUROC (0.59 and 0.60) and lower PPV at the 90th percent cut point (0.08 and 0.07) compared to the published single-center results. We developed and tested a new multicenter logistic regression model with 12 features on dataset (1) and recalibrated this model to dataset (2) [Table 1]. We evaluated multiple models, feature selection routines, and imputation methods. Optimal performance was targeted to AUPRC and PPV based on goal implementation without excessive false alarms.
Conclusion
We externally validated and subsequently developed a new optimal AKI prediction model within two of the largest pediatric critical care datasets. We specifically identified features that are relevant and available at the time of prediction. Ongoing work is applying this model using FHIR resources within a silent implementation to monitor performance.
Performance characteristics of the two developed models
| Total, N | Test Set, N | AKI at 72 hrs, N (%) | AUROC | AUPRC | 50th Percent PPV | 90th Percent PPV | |
| VPS-PN (1) | 59,330 | 14,264 | 1,240 (2.2%) | 0.81 | 0.16 | 0.04 | 0.11 |
| PDC (2) | 130,724 | 26,758 | 3,557 (2.7%) | 0.88 | 0.27 | 0.05 | 0.18 |
Funding
- NIDDK Support