Abstract: Little information exists regarding the determinants of provider compliance with laboratory utilization alerts. Using a large dataset of over 21,000 alerts covering 20 tests, we modeld alert acceptance or bypass using binary logistic regression and random forests. Random forests produced a better overall classifier than logistic regression, but the models shared some common predictive features such as previous test normalcy and time from previous order. Incorporating knowledge gained from these predictive features into alert design may improve compliance in the future.
Learning Objective 1: Compare the power of alert features to predict alert complicance using binary classifier methods.
Jeffrey Szymanski (Presenter)
Washington University in St. Louis
Abraham Qavi, Washington University in St. Louis
Ronald Jackups, Washington University in St. Louis