1 Fast Detection of Risk Signals in Post-marketing Drug Surveillance Using Costs in Claims Data: A Machine Learning Approach

We investigate a machine learning approach to fast detection of risk signals in post-marketing drug surveillance using costs in health care insurance claims data. We show that by employing a locally weighted linear regression model to predict post-drug cost distribution of a population taking a well-known risky pain killer (Vioxx), the safety signal can be discovered four months earlier compared to a recent study using the same datasets. This project demonstrates the potential value of machine learning algorithms in improving real-time post-marketing drug surveillance. Traditional post-marketing drug surveillance systems using health care insurance claim data monitor procedure codes and diagnoses codes to detect adverse drug events (ADEs) [1,2,4,5]. One recent study [3] designed an active post-marketing drug surveillance system from a new angle−monitoring costs in health care insurance claims data, comparing post-treatment cost profiles of populations using alternative drugs based on risk-adjusted group-sequential analysis, and detecting increased spendin