EHR-derived phenotype risk score shows potential to identify lupus patients earlier


An algorithm that sorts through patients’ billing codes in an electronic health record and then assigns a phenotype risk score may be a novel way to identify patients with systemic lupus erythematosus (SLE) earlier, according to research presented at the virtual annual meeting of the American College of Rheumatology.

“Phenotype risk scores measure the degree to which patient’s symptoms, captured by billing codes, overlap with defined disease criteria. Phenotype risk scores may help identify patients earlier with rare diseases,” said presenter and first author April Barnado, MD, assistant professor in the division of rheumatology and immunology and research director of the Lupus Clinic at Vanderbilt University, Nashville, Tenn.

Dr. Barnado and colleagues at Vanderbilt used deidentified data from an EHR with over 3.2 million subjects to find 837 cases of SLE that had been diagnosed by a rheumatologist and 3,149 age-, sex-, and race-matched controls without any autoimmune disease billing codes. They reviewed the charts of all SLE cases, which included both inpatients and outpatients.

Dr. April Barnado
They used ICD-9 and -10 billing codes corresponding to ACR SLE and SLICC disease criteria to build the phenotype risk score. The phenotype risk score was defined as the sum of these codes for each subject weighted by the log inverse prevalence of the code in the entire EHR. They excluded billing codes that specifically mention SLE.

The phenotype risk score was significantly associated with SLE case status after adjusting for age, sex, and race. The mean score among SLE cases was 7.2, compared with 1.7 for controls. African Americans with SLE also had a significantly higher mean score than did Whites with SLE (9.6 vs. 6.4). However, females and males overall had a similar mean risk score (7.1 vs. 7.3), she said.

In SLE cases, phenotype risk scores increased over time, and the risk score could potentially discern future cases of SLE by tracking its rise over time. A small cluster of 57 patients with SLE had phenotype risk scores at the 50th percentile or higher before they had their first SLE billing code, Dr. Barnado noted.

The investigators discovered that control patients with high scores had autoimmune diseases, including two with the highest scores who had SLE according to ACR and SLICC criteria and four who had incomplete or probable SLE. Another eight controls had other autoimmune diseases, including discoid lupus, Crohn’s disease, and inflammatory arthritis, she said.

The two controls with definite SLE initially presented as inpatients and received steroids early in their courses for CNS and renal manifestations, according to Dr. Barnado, “but it wasn’t until much later until both received immunosuppressants (i.e., mycophenolate and cyclophosphamide). The patient with renal manifestations had progression of renal disease and required dialysis.”

Based on this initial research, “we propose that phenotype risk scores could serve as a tool to make earlier SLE diagnoses,” Dr. Barnado said in her presentation.

Rheumatologists react

Many rheumatologists who watched Dr. Barnado’s presentation found the approach to early diagnosis very interesting and posed questions about conducting further research.

ACR president David Karp, MD, PhD, chief of the rheumatic disease division in the department of internal medicine at the University of Texas Southwestern Medical Center, Dallas, wondered whether Dr. Barnado and associates would conduct a prospective study to validate the approach in another EHR or medical center, perhaps having the live EHR trigger investigation and follow-up of potential SLE cases when the phenotype risk score indicated. Dr. Barnado replied that the researchers next plan to overlay a genetic risk score on the EHR data, and can then incorporate data from other U.S. centers that are part of the eMERGE (Electronic Medical Records and Genomics) Network. She said they also hope to “deploy this in real time in the EHR and randomize subjects to this risk score versus standard of care and see how the risk score impacts diagnosis rates and ultimately outcomes.”

S. Sam Lim, MD, professor of medicine at Emory University, Atlanta, asked about adding social determinants data such as zip codes, insurance types, and other relevant information to the risk score. Dr. Barnado said that, while the data are deidentified and will remain so in order to include genetic covariates, she is “brainstorming social determinant data that would be readily available in the EHR,” which could include smoking status, obesity, and education level on some patients.

Image credit (top): CentralITAlliance/Getty Images 
Image credit: metamorworks/Getty Images

The research was funded by grants from the National Institute for Arthritis and Musculoskeletal and Skin Diseases and the National Center for Advancing Translational Sciences. Dr. Barnardo disclosed receiving consulting fees from Nashville Biosciences.