The Medication Ordering Safety System: A Pipeline to Predict Adverse Drug Events at the Time of Ordering

Abstract

Despite widespread adoption of electronic health records (EHRs), medication errors persist. Modern clinical decision support (CDS) systems to prevent medication errors still rely on simple alerts, such as those pertaining to allergies, potential interactions between medications, or duplicative therapy. Machine learning methods can be used to train predictive models that inform more intelligent CDS, but this process is resource-intensive and time-consuming. Here we present the Medication Ordering Safety System (MOSS), a platform we built to catalyze the process of creating predictive models around medication safety. MOSS includes multiple predictive modeling modules, including a cohort creation module, a data extraction module to query the clinical data warehouse, a dataset creation module which joins and prepares the data to be used by modeling packages, and model training and evaluation modules. The ability of MOSS to facilitate adverse event prediction is demonstrated for two inpatient use cases: predicting antihypertensive-induced hypotension and predicting insulin-induced hypoglycemia. The AUC using logistic regression was 0.78 for the hypotension model and 0.74 for the hypoglycemia model. The MOSS pipeline demonstrates an effective platform to facilitate the use of machine learning methods on EHR data to build predictive models around medication safety.