(136) Data-Driven Approaches to Improve Phenotype Sensitivity Using EHR Data

Sunday, August 28, 2022

8:00 AM – 1:45 PM CEST

Location: Hall C

Publication Number: 1181

Presenting Author(s)

Darren Toh, DSc, FISPE
– Harvard Pilgrim Health Care Institute, Boston, Massachusetts, United States

Co-Author(s)

David Carrell, PhD
– Kaiser Permanente Washington Health Research Institute, Seattle, Washington, United States
Joshua C. Smith, PhD
– Assistant Professor, Department of Biomedical Informatics, Vanderbilt University Medical Center, Nashville, TN, United States
Daniel Park
– Vanderbilt University Medical Center, United States
Jill Whitaker
– Vanderbilt University Medical Center, United States
Michael F. McLemore, RN
– Nurse Informatics Data Analyst, Vanderbilt University Medical Center, Department of Biomedical Informatics, United States
Elizabeth E. Hanchrow, RN,MSN
– Informatics Research Nurse, Vanderbilt University Medical Center, Franklin, TN, United States
Dax Westerman
– Vanderbilt University Medical Center, TN, United States
Joshua Osmanski
– IT Project Manager, Vanderbilt University Medical Center, Nashville-Davidson, TN, United States
Robert Winter
– Vanderbilt Health, Nashville, United States
Saranrat Wittayanukorn, PhD
– Epidemiologist, U.S. Food and Drug Administration, Silver Spring, MD, United States
Danijela Stojanovic, PharmD, PhD
– U.S. Food and Drug Administration, Silver Spring, MD, United States
Arvind Ramaprasan
– Data Reporting & Analytics Consultant, Kaiser Permanente Washington Health Research Institute, United States
Ann Kelley, MHA
– Research Specialist II, Kaiser Permanente Health Plan of Washington, United States
Mary Shea, M.A.
– Manager, Research Health Information, Kaiser Permanente Washington Health Research Institute, WA, United States
David J. Cronkite, MS
– Computational Linguist, Kaiser Permanente Washington Health Research Institute, United States
Jing Zhou, PhD
– Kaiser Permanente Washington, United States
Yueqin Zhao, PhD
– US Food and Drug Administration, United States
Kevin B. Johnson, MD, MS
– David L. Cohen University Professor, Perelman School of Medicine at the University of Pennsylvania, philadelphia, PA, United States
David Aronoff
– Indiana University School of Medicine, United States

Background: Medical product safety studies traditionally use condition-specific diagnosis codes as filters to identify patients with health outcomes of interest, but such filters may lack sensitivity.

Objectives: We sought to identify five types of electronic health record (EHR) data (coded procedures, labs, medications, problem lists, and diagnoses; hereafter called features) as surrogates for diagnosis codes used in traditional filters. With COVID-19 disease as a use case, we evaluated whether such surrogates improved sensitivity by identifying COVID-19 cases overlooked by a traditional filter.

Methods: Using EHR data from Vanderbilt University Medical Center (VUMC) and Kaiser Permanente Washington (KPWA) from 4/2020-3/2021 we identified a cohort of potential COVID-19 cases using six COVID-19-specific diagnosis codes as a traditional filter. We noted all other EHR features appearing in close proximity to their COVID-19-coded encounters. For each such feature we calculated a metric indicating how much more often the feature co-occurred with COVID-19-coded encounters compared to encounters of patients never receiving COVID-19 diagnoses. For example, the extracorporeal membrane oxygenation (ECMO) procedure was 40.8 times more likely in COVID-19-coded VUMC encounters. A clinical domain expert reviewed all features that were at least 10 times more common in COVID-19 encounters and retained those with clinical face validity. We used retained features to create five “high-sensitivity” filters, one per data type, and applied them to all patients not identified by the traditional filter. We used manual chart review of random samples to estimate true cases of COVID-19 with at least moderate symptoms and, separately, at least mild symptoms.

Results: The traditional filter identified 20,951 VUMC and 6,847 KPWA potential cases; the five high-sensitivity filters added 4,566 (VUMC, +21.8%) and 1,482 (KPWA, +21.6%) potential cases. Estimated true case rates for potential cases at VUMC by the traditional filter were 51% and 90% (respectively, for at least moderate and at least mild symptom severity); corresponding rates at KPWA were 34% and 71%. Estimated true-case rates for potential cases by the high-sensitivity filter at VUMC were 6.4% and 55% (respectively, for at least moderate and at least mild symptoms); corresponding true case rates at KPWA were 13% and 38%. Importantly, the high-sensitivity filters increased true case sensitivity at VUMC by 2.7% and 13.2% (moderate, mild) and at KPWA by 7.9% and 9.8% (moderate, mild).

Conclusions: This high-sensitivity filter approach can improve the sensitivity of traditional filters, is largely automated, and may generalize to other health outcomes.