Study Design and Administration
The remote monitoring program was implemented as a quality improvement project at BJC Healthcare and Washington University School of Medicine, a large university-affiliated healthcare system based in St. Louis, Missouri. A multidisciplinary team of clinicians, including internal medicine, pulmonology, and infectious disease specialists, refined the care pathway used in the program. This pathway was modeled on the program developed by the Cleveland Clinic using Epic’s Care Companion Home Monitoring Program app.17,18 In parallel, an operational team allocated resources for patient monitoring and configured the Care Companion platform to accommodate the desired workflows in conjunction with analysts from Epic electronic health record (EHR). The first patient was enrolled on 4/6/2020. This manuscript presents a retrospective propensity-matched cohort study of patients who enrolled in the home monitoring program and COVID-positive outpatients who did not enroll in the program. The Human Research Protection Office at Washington University School of Medicine approved this cohort study (#202,103,247) and granted a waiver of informed consent to obtain data from the EHR. This report was written in compliance with STROBE guidelines.19
This study includes all patients who enrolled in the home monitoring program between 4/6/2020 and 12/7/2020. Patients were enrolled either by outreach from program staff or by their healthcare provider offering the program to them. Inclusion criteria for outreach from program staff included (1) a positive COVID-19 lab result in the past 14 days, (2) “confirmed COVID-19 infection” added to the EHR problem list within the past 30 days, or (3) an active COVID-19 isolation flag in the EHR. Exclusion criteria included age < 18 or inability to converse in English. Eligible patients with active Epic MyChart accounts received automated invitations in English to enroll through the app. Eligible patients without active MyChart accounts or who did not respond were enrolled by telephone.
Matched controls were selected from the group of adult patients residing in Missouri or Illinois with a positive COVID test performed in an outpatient setting in the BJC network between 4/1/2020 and 12/7/2020. Because the controls were limited to outpatients, any enrolled patients whose initial COVID test had been performed in an inpatient setting were excluded from the propensity-matched analyses.
Each morning, patients completed a structured questionnaire via the Care Companion app or via telephone assessing appetite, cough, diarrhea, fever, shortness of breath, vomiting, and weakness (Supplement Figure S1). Temperature and oxygen saturation were also reported if the patient could measure them. Patients could choose between the app and telephone calls at the time of enrollment. Patients in the app arm who did not complete the questionnaire received an outbound call from a medical assistant to collect responses. If the patient’s symptoms or self-reported vital signs fell outside pre-specified parameters (Supplement Table S1), then the case was escalated to a triage nurse. The nurse phoned the patient the same day to assess further and either advised the patient to self-monitor, arranged a video visit with a physician or an in-person respiratory clinic visit, or instructed the patient to seek emergency medical care. Daily symptom monitoring continued until symptom resolution, or for 14 days if the patient remained asymptomatic. Patients were automatically removed from the program if they were admitted to a hospital or if they could not be contacted for three consecutive days. In addition, patients could opt out at any time. Daily symptom questionnaire responses and notes documenting all phone calls were stored in the EHR.
Among the enrollees, key exposure variables included race and neighborhood disadvantage. Race was extracted from the EHR. The degree of neighborhood disadvantage was estimated using the Area Deprivation Index (ADI) score associated with the patients’ zip codes.20 Patients residing in zip codes with an ADI score greater than the 6th nationwide decile (which was the median value in this dataset) were classified as experiencing neighborhood disadvantage.
Among enrollees, the primary outcome metric compared across race and neighborhood disadvantage groups was retention in the program. Retention was defined as the percentage of enrollees who remained in the program until symptom resolution (or until 14 days if asymptomatic) or until hospital admission.
Among enrollees and matched controls, the outcome metrics were emergency department visits and hospital admissions. Encounters at any of the 12 BJC locations in Missouri or Illinois that use the Epic EHR were included. Each ED visit or admission was classified as potentially COVID-related if the primary diagnosis and/or primary problem associated with encounter contained any of the following text strings: anosmia, asthma, bronch*, COPD, corona*, cough, COVID, dyspnea, emesis, fatigue, febrile, fever, flu, hypox*, leukocytosis, mental status, myalgia, nausea, pneumonia, pulmonary, respiratory, sepsis, septic, shortness of breath, smell, SOA, SOB, URI, viral, virus, vomit, weak. For program enrollees, each ED visit or admission was further classified by whether the patient’s case had been escalated to a home monitoring triage nurse during program enrollment.
Analyses were conducted in R version 4.0.03 (R, Vienna, Austria). All available data were used, so the sample size was determined by the number of patients who had enrolled in the home monitoring program (with no sample size calculation). Demographic characteristics of the population were compared across race and neighborhood disadvantage strata using Pearson’s chi-squared test (categorical variables), Fisher’s exact test (race), or the Wilcoxon rank sum test (age). Patient zip codes were used to classify each patient as living either in the urban core of the city (Rural–Urban Commuting Area (RUCA) code equal to 1) or not.21 Program arm selection (telephone versus app) and retention were compared across race and neighborhood disadvantage strata using a chi-squared test. Comparisons across race strata were limited to White patients and Black patients due to the small number of patients reporting other races.
For the matched analysis, enrollees (excluding those initially diagnosed with COVID as an inpatient) were matched 1:2 to control patients based on the propensity score for enrollment. Propensity scores were calculated using a logistic regression adjusting for age, sex, ethnicity, residence in the urban core, diabetes, hypertension, coronary artery disease, heart failure, arrhythmia, chronic lung disease, chronic kidney disease, chronic liver disease, hypothyroidism, depression, and any cancer. Satisfactory matches had propensity scores within a caliper of 0.1 standard deviation, plus exact matches for race and neighborhood disadvantage. Matching was performed using the MatchIt package.22 Time from positive coronavirus test to ED visit and to hospital admission was compared between the matched groups in a Cox proportional hazards model. Patients were right-censored in the ED visit analysis if they were directly admitted to a hospital. The Cox model was repeated within each race and neighborhood disadvantage stratum, and within each program arm. Statistical significance was determined using α = 0.05.