A Machine Learning-Based Approach to Predicting Acute Kidney Injury and Associated Medication Regimen Use in Critically Ill Adults

Speaker(s)

Brothers T1, Al-Mamun M2
1University of Rhode Island, Kingston, RI, USA, 2West Virginia University, Morgantown, WV, USA

OBJECTIVES: Our primary aim was to evaluate the performance of several machine learning models to predict AKI in the ICU using MRCI scores.

METHODS: A retrospective observational study of 322 critically ill adults from February 1 to August 30, 2020, was performed. Excluded patients were long-term care patients, history of renal replacement or kidney transplantation. AKI status was determined according to the 2012 Kidney Disease Improving Global Outcomes guidelines. Descriptive statistics were conducted using an independent t-test, chi-square test, or fisher’s exact test. Predictors of interest included MRCI scores at time of hospitalization, evaluation of 14 medication classes, and demographics. Machine learning algorithms used were logistic classifier, random forest, and XGBoost. Predictors of interest were ranked by variable importance to aid in the prediction of AKI.

RESULTS: Among 319 patients, 153 (48%) met AKI criteria with a mean age of 62.54± 17.3 years. MRCI scores were found to be statistically significantly (p <0.001) between AKI and non-AKI individuals (mean 42.11± 36.15 vs. 33.09± 31.7 and 76.47± 48.1 vs. 59.02±32.60), respectively. Patients with AKI had acute respiratory failure (59%), acidosis (53%), and hypokalemia (53%). AKI patients commonly received endocrine agents (51.4%), pulmonary agents (53%), and vasopressors (59%) on day 1 of ICU admission. The XGBoost ML model had the highest prediction accuracy of AKI (74%; IQR: 62-83) with an average sensitivity of 68% compared to other ML models. Top variables of importance in predicting AKI in the ICU were MRCI scores at 24 hours, MRCI scores of outpatient medication use, BMI, age, and acute respiratory failure.

CONCLUSIONS: The findings of our study accurately predicted AKI in the ICU setting 74% of the time. Thereby, suggesting that incorporation of an MRCI score into the clinical decision-making process may aid clinicians in the early identification of at-risk patients to proactively implement preventative care strategies.

Code

PT16

Topic

Clinical Outcomes, Methodological & Statistical Research

Topic Subcategory

Artificial Intelligence, Machine Learning, Predictive Analytics, Clinical Outcomes Assessment, Performance-based Outcomes

Disease

Drugs, Personalized & Precision Medicine, Urinary/Kidney Disorders