Abstract: FR-PO082
Predicting Persistent AKI Using Machine Learning: A Multicenter External Validation Study
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
- Kashani, Kianoush, Mayo Clinic Minnesota, Rochester, Minnesota, United States
- Zappalà, Simone, U-Care Medical srl, Torino, Italy
- Alfieri, Francesca, U-Care Medical srl, Torino, Italy
- Ancona, Andrea, U-Care Medical srl, Torino, Italy
Background
Acute Kidney Injury (AKI) correlates with higher morbidity and mortality, and its longer duration leads to acute and chronic kidney disease progression. Prompt interventions with externally validated prediction models for those at increased risk of AKI could change its course. We aim to validate the Persistent Electronic Alert (PersEA), a machine learning model, using routinely collected medical data in intensive care units, in predicting Persistent AKI.
Methods
Acute Kidney Injury (AKI) was defined and staged using Kidney Disease Improving Global Outcomes guidelines. Persistent AKI was defined as AKI stage 3 lasting for ≥72 hours or leading to death or necessitating renal replacement therapy. This retrospective study included Adult (≥18 years old) admissions at the Mayo Clinic exhibiting at least AKI stage 2. Model performance was assessed using a specific metric that penalizes late alarms, measuring the area under the receiver operating characteristic (auROC) and the area under the Precision-Recall (auPR) curves. The generalizability of the alerting system was measured through sensitivity, specificity, precision, and lead time under different cutoff criteria. The dataset was stratified into subpopulations according to the reason for admission, comorbidities, and demographics to measure possible biases.
Results
Among the ICU admissions at the Mayo Clinic, a cohort comprising 5,589 cases met the selection criteria and were subsequently included in the conclusive analyses. While the internal validation cohort demonstrated an 11% incidence of persistent AKI, the Mayo Clinic cohort exhibited a lower incidence of 5%. The PersEA model achieved an auROC of 0.98 (95% CI, 0.97-0.98) and an auPR of 0.67 (95% CI, 0.60-0.73). By selecting the threshold that reached 0.80 sensitivity in the internal cohort, PersEA achieved 0.88 sensitivity, 0.94 specificity, and 0.47 precision on Mayo Clinic data.
Conclusion
PersEA model exceptionally performed on an external cohort, showing that the model is scalable on high-quality data with little to no tuning once a noisy training set is chosen.
Funding
- Commercial Support – U-Care Medical srl