P11: Applying Natural Language Processing to the Extraction of Medical Safety Information from Free-Text on the CIOMS Form
Bristol-Myers Squibb Company United States
The objective was to develop a fit-for-use, proprietary, web-based tool built on natural language processing (NLP) to automatically extract information from free-text medical records of Council for International Organizations of Medical Sciences (CIOMS) forms.
The NLP web-based tool was developed for a single use case to extract medications, conditions, and procedures presented in the “Describe Reaction(s)” section of the CIOMS form. The NLP-based algorithm identifies several entities and their relationship with others within text.
A use case will be presented in which the NLP algorithm was used to search CIOMS forms for potential confounding medications and potential drug-drug interactions among patients who had interstitial nephritis during atazanavir/cobicistat intervention. The algorithm was also tested on 3 additional use cases with a total of 189 CIOMS forms to determine its performance. Overall sensitivity was 80% and specificity was 98%.
Nearly 80% of healthcare data are unstructured and available in various sources. Data from the adverse reaction description field on a CIOMS form, used for the reporting of suspected adverse reactions to any particular medical product, is one type of the unstructured text data.
Usually, when a safety concern arises, pharmacovigilance (PV) scientists and medical safety assessment physicians (MSAPs) need to manually review dozens or hundreds of CIOMS forms to assess causality between reported adverse events and a suspected product of interest. The task requires significant time, effort, and resources.
The system-agnostic tool we are developing could automatically scan thousands of CIOMS forms and extract all (or specific) medications and conditions within hours. Use of the NLP algorithm could result in considerable time savings, thus improving the efficiency of CIOMS safety review during medical safety signal monitoring and assessment. In addition, the extracted data could be exported in a format that is structured to support summary statistics and further analysis.