Developing an expert panel process to refine health outcome definitions in observational data

Objectives: Drug safety surveillance using observational data requires valid adverse event, or health outcome of interest (HOI) measurement. The objectives of this study were to develop a method to review HOI definitions in claims databases using (1) web-based digital tools to present de-identified patient data, (2) a systematic expert panel review process, and (3) a data collection process enabling analysis of concepts-of-interest that influence panelists' determination of HOI.

Methods: De-identified patient data were presented via an interactive web-based dashboard to enable case review and determine if specific HOIs were present or absent. Criteria for determining HOIs and their severity were provided to each panelist. Using a modified Delphi method, six panelist pairs independently reviewed approximately 200 cases across each of three HOIs (acute liver injury, acute kidney injury, and acute myocardial infarction) such that panelist pairs independently reviewed the same cases. Panelists completed an assessment within the dashboard for each case that included their assessment of the presence or absence of the HOI, HOI severity (if present), and data contributing to their decision. Discrepancies within panelist pairs were resolved during a consensus process.

Results: Dashboard development was iterative, focusing on data presentation and recording panelists' assessments. Panelists reported quickly learning how to use the dashboard. The assessment module was used consistently. The dashboard was reliable, enabling an efficient review process for panelists. Modifications were made to the dashboard and review process when necessary to facilitate case review. Our methods should be applied to other health outcomes of interest to further refine the dashboard and case review process.

Conclusion: The expert review process was effective and was supported by the web-based dashboard. Our methods for case review and classification can be applied to future methods for case identification in observational data sources.