Comparing Machine Learning Algorithms for Predicting Acute Kidney Injury. Issue 7 (July 2019)

• Record Type: Journal Article
• Authors: Parreco, Joshua; Soe-Lin, Hahn; Parks, Jonathan J.; Byerly, Saskya; Chatoor, Matthew; Buicko, Jessica L.; Namias, Nicholas; Rattan, Rishi
• Is Part Of: American surgeon. Volume 85:Issue 7(2019)
• Page Start: 725
• Page End: 729
• Publication Date: 2019-07
• Subjects: Surgery -- Periodicals; Surgery -- United States -- Periodicals; 617.0973
• DOI: 10.1177/000313481908500731
• Languages: English
• ISSNs: 0003-1348

Abstract

Prior studies have used vital signs and laboratory measurements with conventional modeling techniques to predict acute kidney injury (AKI). The purpose of this study was to use the trend in vital signs and laboratory measurements with machine learning algorithms for predicting AKI in ICU patients. The eICU Collaborative Research Database was queried for five consecutive days of laboratory measurements per patient. Patients with AKI were identified and trends in vital signs and laboratory values were determined by calculating the slope of the least-squares-fit linear equation using three days for each value. Different machine learning classifiers (gradient boosted trees [GBT], logistic regression, and deep learning) were trained to predict AKI using the laboratory values, vital signs, and slopes. There were 151, 098 ICU stays identified and the rate of AKI was 5.6 per cent. The best performing algorithm was GBT with an AUC of 0.834 ± 0.006 and an F -measure of 42.96 per cent ± 1.26 per cent. Logistic regression performed with an AUC of 0.827 ± 0.004 and an F -measure of 28.29 per cent ± 1.01 per cent. Deep learning performed with an AUC of 0.817 ± 0.005 and an F -measure of 42.89 per cent ± 0.91 per cent. The most important variable for GBT was the slope of the minimum creatinine (30.32%). This study identifies the best performing machine learning algorithms for predicting AKI using trends in laboratory values in ICU patients. Early identification of these patients using readily available data indicates that incorporating machine learning predictive models into electronic medical record systems is an inevitable requisite for improving patient outcomes.