Detecting possible vaccination reactions in clinical notes.

Abstract

The Vaccine Safety Datalink is a collaboration between the CDC and eight large HMO's to investigate adverse events following immunization through analysis of medical care databases and patients' medical charts. We modified an existing system called MediClass that uses natural language processing (NLP) and knowledge-based methods to classify clinical encounters recorded in electronic medical records (EMRs). We developed the knowledge necessary for MediClass to detect possible vaccine reactions in the outpatient, ED, and telephone encounters recorded in the EMR of a large HMO. We first trained the system using a manually coded gold standard training set, and achieved high sensitivity and specificity. We then ran a large set of post-immunization encounter records through MediClass to see if our method would generalize. Compared to methods that use administrative and clinical codes assigned to the EMR by clinicians, the system significantly improves the positive predictive value for detecting possible vaccine reactions.