Abstract: FR-PO708
Automated Learning and Early Recognition Technology for Neonatal AKI (ALERT-NAKI)
Session Information
- Pediatric Nephrology - 1
October 25, 2024 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Pediatric Nephrology
- 1900 Pediatric Nephrology
Authors
- Mohamed, Tahagod, Nationwide Children's Hospital, Columbus, Ohio, United States
- Slaughter, Jonathan L., Nationwide Children's Hospital, Columbus, Ohio, United States
- Bambach, Sven, Nationwide Children's Hospital, Columbus, Ohio, United States
- Magers, Jacqueline K., Nationwide Children's Hospital, Columbus, Ohio, United States
- Rust, Laura, Nationwide Children's Hospital, Columbus, Ohio, United States
- Patel, Shama, Nationwide Children's Hospital, Columbus, Ohio, United States
- Rust, Steve, Nationwide Children's Hospital, Columbus, Ohio, United States
- Spencer, John David, Nationwide Children's Hospital, Columbus, Ohio, United States
- Wilson, Francis Perry, Yale University, New Haven, Connecticut, United States
Background
Neonatal acute kidney injury (AKI) is underrecognized despite its high prevalence and known adverse effects on short- and long-term outcomes. Evidence-based tools for providers to predict and recognize neonatal AKI risk are limited. Under recognition of neonatal AKI prohibits providers from timely interventions on modifiable AKI risk factors. We hypothesized that neonatal AKI occurrence can be predicted by utilizing existing EHR data and machine learning algorithms.
Methods
Using retrospective EHR data from 5400 infants admitted in 2017-2021 to our level IV NICU for >183,000 patient days, we trained a machine learning model to predict neonatal AKI occurrence. Data were collected after local IRB approval. Known neonatal AKI risk factors were identified from existing literature. Neonatal AKI was defined by changes in serum creatinine (sCr) and urine output (UOP) according to the neonatal modification of KDIGO criteria. We used least absolute shrinkage and selection operator (LASSO) algorithm to identify reliable neonatal AKI predictors and to develop a predictive model. The area under the curve (AUC) was utilized to evaluate model performance
Results
When provided with 34 known neonatal AKI risk factors, LASSO chose 8 reliable predictors including fluid balance status, hypotension requiring vasopressors, invasive ventilation, FiO2%, congenital anomalies of the kidneys and urinary tract, sepsis, sCr monitoring and value changes. The model predicts significant changes in sCr and UOP during the next 48 hours with an AUC of 0.76, Figure 1.
Conclusion
Neonatal AKI can be reliably predicted by utilizing machine learning algorithms and available EHR data. This is the first step to empower providers with evidence-based tools to improve neonatal AKI prediction and recognition and to allow time for intervention on modifiable risk factors before kidney injury occurs.
Funding
- Private Foundation Support