Predicting Acute Kidney Injury with Nephrotoxic Burden in Noncritical Patients: An Internal and External Validation...

Area of Science:

• Nephrology
• Artificial Intelligence
• Medical Informatics

Background:

• Acute kidney injury (AKI) affects approximately 20% of hospitalized patients.
• AKI can lead to irreversible kidney damage if not managed promptly.
• Early prediction of AKI progression is crucial for timely intervention.

Purpose of the Study:

• To develop and validate deep learning models for predicting AKI progression.
• To identify key predictors of AKI progression within 48 hours.
• To assess the performance of models trained on multi-center data.

Main Methods:

• Retrospective study using electronic health records from two large healthcare systems (UPMC and UFH).
• Development of deep learning models utilizing demographics, comorbidities, medications, lab results, and vital signs.
• External validation of models to predict Stage 2 or higher AKI based on KDIGO serum creatinine criteria.

Main Results:

• Models demonstrated robust predictive performance, with Area Under the Receiver Operating Characteristic Curve (AUROC) values ranging from 0.77 to 0.84.
• The model trained on multi-center data (UFH-UPMC Model) achieved AUROC of 0.81 for UFH and 0.82 for UPMC test cohorts.
• Key predictors included kinetic estimated glomerular filtration rate, nephrotoxic drug burden, and blood urea nitrogen.

Conclusions:

• A multi-center deep learning model shows strong performance in predicting AKI progression.
• Implementation of such models could aid in the prevention of AKI progression.
• The findings highlight the potential of AI in improving acute kidney injury management.