Ascertainment of Stroke from Administrative Data to Support a Pragmatic Embedded Clinical Trial

The goal of this observational study was to identify stroke hospitalizations using International Classification of Disease (ICD)-10 codes and use these codes to develop an ascertainment algorithm for use in pragmatic clinical trials, reducing or eliminating the need for manual chart adjudication in future. Using VA (Veterans Affairs) electronic medical records, 9959 patient charts with ICD-10 codes indicating stroke were screened and a sample of 304 were adjudicated by three clinical reviewers. Hospitalizations were categorized as stroke or non-stroke and positive predictive value (PPV) was calculated for each ICD-10 code that was sampled. The adjudicated codes were categorized for use in a decision tool for identifying stroke in a clinical trial. Of the 304 hospitalizations adjudicated, 192 met the definition of stroke. Of the ICD-10 codes evaluated, I61 yielded the highest PPV (100%) while I63.x yielded the 2nd highest PPV (90%) with a false discovery rate of 10%. A relatively high PPV of ≥80% was associated with codes I60.1-7, I61, I62.9 and I63, which accounted for nearly half of all cases reviewed. Hospitalizations associated with these codes were categorized at positive stroke cases. The incorporation of large administrative datasets, and elimination of trial specific data collection, increases efficiencies, while reducing costs. Accurate algorithms must be developed to allow for identification of clinical endpoints from administrative databases to offer a reliable alternative to study-specific case report form completion. This study demonstrates an example of how to apply medical record data to a decision tool for clinical trial outcomes. CSP597 or clinicaltrials.gov NCT02185417.