Predicting Nephrotoxic Acute Kidney Injury in Hospitalized Adults: A Machine Learning Algorithm

Area of Science:

• Nephrology
• Artificial Intelligence
• Clinical Informatics

Background:

• Acute kidney injury (AKI) is a frequent complication in hospitalized adults, posing significant clinical challenges for prediction and prevention.
• Existing predictive models often struggle with accuracy in adult populations, necessitating advanced approaches.

Purpose of the Study:

• To adapt and enhance the pediatric Nephrotoxic Injury Negated by Just-in Time Action (NINJA) program using machine learning for improved prediction of nephrotoxic AKI in adults.
• To evaluate the performance of a recurrent neural network (RNN) gated recurrent unit (GRU) model in identifying adult patients at high risk for AKI following nephrotoxic exposure.

Main Methods:

• A retrospective cohort study involving adult patients admitted to the University of Iowa Hospital between 2017 and 2022.
• Data included 85 variables: demographics, laboratory tests, vital signs, and medications. AKI was defined by a serum creatinine increase of ≥0.3 mg/dL.
• A RNN-GRU model was trained on 85% of the data and tested on the remaining 15%, with performance assessed using precision, recall, negative predictive value, and AUC.

Main Results:

• The study analyzed 14,480 patients and 37,300 high-nephrotoxin exposure events; 29% of exposures led to AKI within 48 hours.
• The RNN-GRU model achieved a precision of 0.60 for AKI prediction, reducing false alerts from 2.5 to 0.7 per AKI case.
• Key risk factors identified included lowest hemoglobin, lowest blood pressure, and highest white blood cell count, along with specific medications like acyclovir and piperacillin-tazobactam.

Conclusions:

• The developed RNN-GRU model significantly decreases false alerts for nephrotoxic AKI in adults.
• This advancement supports the potential translation of the NINJA program to adult healthcare settings, facilitating more precise interventions.
• Further research should incorporate comprehensive clinical variables and drug-specific data for enhanced predictive accuracy.