SW-3B - Developing Computable Phenotypes for Medical Product Safety Surveillance: Innovative Methods and Data Sources

Background: Accurate phenotyping of health outcomes is crucial for medical product safety surveillance and robust pharmacoepidemiologic studies. Developing and validating outcome phenotypes can be laborious and costly, and some outcomes are difficult to measure accurately in claims data, which are commonly used for such studies. Innovative methods are needed to support scalable development of highly accurate phenotypes.

Objectives: To present innovative methods and approaches to support computable phenotype development using electronic health data. This symposium will benefit regulators, researchers and industry partners seeking to improve ability to measure safety outcomes from routinely available health data.

Description: First, for context, we will describe the vision for integrating computable phenotyping and electronic health records (EHR) data into US surveillance efforts, based on the perspective of leaders from the FDA Sentinel Innovation Center. Next, Dr. Sascha Dublin will present results from a Sentinel project that developed processes and tools for assessing the difficulty of generating a computable phenotype using EHR data, including Natural Language Processing (NLP), for outcomes that are challenging to measure from claims data. Dr. Robert Penfold will present experiences with using NLP in a distributed data setting to create new variables for the Sentinel Common Data Model derived from free text notes in EHRs, including measures for smoking status, obesity, and anaphylaxis. Dr. Joshua Smith will present work using PheNorm, an automated process to guide the development of computable phenotypes using NLP to make phenotype development faster and less resource-intensive. Dr. Richeek Pradhan will present research about PhenoScale, a semi-automated, data-driven workflow for phenotype development that identifies “features” from claims data using risk set sampling and a LASSO model. He will present results from applying this workflow to acute kidney injury, resulting in a final model that achieved superior performance (AUROC 0.92) compared with expert-curated codes (AUROC 0.77). Dr. Albert Prats-Uribe will present work done by DARWIN EU® to develop a process and accompanying tools for standardized, reproducible phenotype development and storage using the OMOP Common Data Model. Presentations will include practical advice, tools, and resources to enable participants to apply these methods in their own context. We will conclude with a panel discussion including US and EU regulators about future applications of these methods and directions for new research to support phenotype development.