3165 — Ascertainment of Clinical Outcomes from Electronic Medical Record Data for Point-of-Care Clinical Trials
Leatherman SM, Boston VA Healthcare; Riley KE, Boston VA Healthcare; Woods PA, Boston VA Healthcare; Zimolzak AJ, Boston VA Healthcare; Majahalme N, Boston VA Healthcare; Kudesia V, Boston VA Healthcare; Ferguson RE, Boston VA Healthcare; Fiore LD, Boston VA Healthcare;
Point-of-care (POC) clinical trials are pragmatic trials designed to reduce the burden of research by embedding the study into routine clinical care. After enrollment and randomization, patients are passively followed through the medical record for study outcomes. Thus, the success of POC clinical trials depends on the ability to assess outcomes directly from the medical record. Unfortunately, electronic medical record (EMR) data are not collected with research in mind. And, for many clinical outcomes, this assessment cannot be done with cleanly structured data in the EMR.
We use the development of an algorithm to predict acute decompensated heart failure for the VA Diuretic Comparison Project as an example of an approach to outcome ascertainment using EMR data. Patients at a single VA medical center were identified for hospitalizations possibly associated with heart failure by a collection of 33 potentially relevant ICD9 codes. Hospitalizations were independently adjudicated by two residents. The residents and study team together established a set of characteristics used to determine whether a patient was hospitalized for heart failure. These elements were located in the data warehouse associated with the EMR, curated, cleaned, and codified for statistical modelling.
Residents evaluated 451 hospitalizations between 2009 and 2014 and assigned diagnoses as either likely heart failure or likely not heart failure. Of these, 399 hospitalizations received the same diagnosis from both doctors, 347 (87%) for confirmed heart failure. We identified six structured indicators of heart failure (discharge ICD9 code, BNP, weight change, various medications received during hospitalization, and implantable cardioverter defibrillator placement) and nine key phrases present in physicians' notes most often used to make a diagnosis. Text parsing was used to search admission and discharge notes for presence of these keywords including 'diuresis', 'resolution of symptoms', 'volume overload', and 'euvolemia'.
Logistic regression and regression trees were used to improve prediction of heart failure done by ICD9 code alone.
Although the development of outcome prediction models requires a significant time investment in chart review, the ultimate goal of these activities is to create a reusable catalogue of curated events for future POC studies, saving resources in the long run.