|Friday, February 16|
|PS2 Poster Session 2 and Refreshments||
Fri, Feb 16, 5:15 PM - 6:30 PM
Consensus Strategy for Variable Selection in Clinical Prediction Rule Development (303631)
Jodi Lapidus, OHSU/PSU School of Public Health
Jessina C. McGregor, OHSU/OSU College of Pharamcy
Keywords: prediction rule, consensus strategy, multidimensional data
Data-driven tools to inform clinician antibiotic prescribing would greatly support efforts to control antibiotic resistance. We outline a consensus strategy to develop a prediction rule to direct appropriate selection of antibiotic agents to treat urinary tract infections. Random forest, adaptive LASSO, and boosted classification tree approaches were applied to multidimensional, electronic health record data to robustly and efficiently reduce candidate predictors for a multivariable logistic regression. Variables deemed significant by any method were included in subsequent best subsets model building. The resulting classifier was evaluated using area under the curve (AUC) of the receiver operating characteristic curve (ROC), sensitivity, and specificity. A validation dataset was set aside to independently validate prediction rule performance, but the rule did not attain a priori minimum acceptable sensitivity and specificity thresholds set by clinical partners on the development dataset. Plans to augment data as well as refine the model will be implemented under existing consensus framework. Data management was conducted in SAS v9.4 and statistical analyses in R 3.3.3.