Relative Importance of Predictors of Artificial Neural Network Modeling Results with Applications to Evaluating Vasopressor Treatments for Subarachnoid Hemorrhage (SAH) Patients

• Activity Number: 618 - Machine Learning for Big Data
• Type: Contributed
• Date/Time: Thursday, August 1, 2019 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Duo Yu* and Hulin MI Wu
• Companies: University of Texas Health Science Center at Houston and University of Texas Health Science Center at Houston
• Keywords: EHR; MLP; relative importance of predictors; interpretability of artificial neural network models

Abstract

Subarachnoid Hemorrhage (SAH), bleeding into the subarachnoid space surrounding the brain. Three types of vasopressors are commonly used for SAH patients. However, there is no clear guidance on which vasopressor should be assigned to each individual patient. In this study, we used the multilayer perceptron (MLP) to predict the mortality of SAH patients under different vasopressor treatments based on a large EHR database, Cerner Health Facts® Database. From the Cerner EHR database with 49.5 million patients, we identified 4645 patients with SAH under at least one of the three vasopressor treatments. The potential predictors for the mortality include comorbidities of disease diagnosis, medications, demographic variables and the vasopressor use with a total of 919 variables. Our ANN model reaches 96% AUC. We also proposed a novel approach to interpreting the results of the ANN modeling. We defined a new “relative importance” index based on the positive and negative relative risk for each predictor in the ANN model. We compared our new relative importance measure with the existing methods. Our ANN modeling also suggests that 97% of patients should be assigned to phenylephrine.