Predicting acute kidney injury in critically ill patients using comorbid conditions utilizing machine learning

Area of Science:

• Nephrology
• Critical Care Medicine
• Data Science

Background:

• Acute kidney injury (AKI) is a growing concern in intensive care units (ICUs), associated with poor patient outcomes.
• Early identification of patients at risk for AKI is crucial for timely intervention and improved prognosis.

Purpose of the Study:

• To develop and validate a predictive model for AKI using pre-ICU admission data.
• To leverage machine learning for early AKI risk stratification in ICU settings.

Main Methods:

• Utilized a large dataset of 98,472 adult ICU admissions from Mayo Clinic and 51,801 encounters from the MIMIC-III cohort.
• Trained a gradient-boosting model on 80% of the Mayo Clinic data, incorporating pre-admission patient features to predict in-ICU AKI.

Main Results:

• AKI occurred in 39.9% of the Mayo Clinic cohort; AKI patients were older with higher mortality.
• The 30-feature predictive model achieved an AUC of 0.690 in the Mayo Clinic cohort and 0.656 in the MIMIC-III cohort.
• The model demonstrated good performance in predicting AKI using only pre-admission data.

Conclusions:

• Machine learning enables prediction of AKI in ICU patients based on pre-admission information.
• The developed model is independent of in-ICU data, offering value for initial patient risk stratification upon admission.