Partners HealthCare is an integrated health delivery system in Eastern Massachusetts. Among other institutions, Partners HealthCare includes Massachusetts General Hospital and Brigham and Women’s Hospital, academic medical centers each with multiple community- and hospital-based primary care clinics.
We obtained electronic health record data for all patients in the exposure and usual care groups. This included sociodemographic variables (gender, age, language preference, partner status, veteran status, race/ethnicity, BMI, health insurance, and medical conditions), systolic blood pressure (SBP) measurements, and healthcare utilization within the partners system (visits to primary care, specialists, and the emergency department; inpatient admissions).
Exposure: Virtual Visits
This virtual visit platform has been described in detail elsewhere.8 In brief, virtual visits are asynchronous, structured digital exchanges between a patient and a primary care clinician following an in-person encounter. Their purpose is to substitute for a follow-up in-person visit. We assessed virtual visits specifically crafted for hypertension follow-up after an in-person visit. After a visit where hypertension was discussed, clinicians could ask their patients to follow-up online through the virtual visit platform from 21 to 180 days following the initial encounter. The follow-up period from in-person visit to virtual visit was entirely at the discretion of the treating clinician. The virtual visit begins with a notification e-mail which directs the patient to a secure mobile-friendly website. The patient enters up to five blood pressure readings, notes medication adherence through a single binary response, relays medication side effects by free text, and can ask the ordering clinician questions. Once submitted, the clinician reviews the data and can respond with structured treatment decisions via the platform, recommendation for repeat virtual visit, recommendation for follow-up phone call, or request for an in-person appointment. A clinician could offer a virtual visit as s/he saw fit. There was no protocolized standardization of the blood pressure device, calibration, or training.
Virtual visits were first offered to select primary care practices at Massachusetts General Hospital in March 2013. In March 2015, virtual visits were made available to all primary care clinicians at Massachusetts General Hospital and its affiliated clinics. Virtual visits have not been available outside of Massachusetts General Hospital and its affiliated clinics, but they are planned to be rolled out to all of Partners HealthCare.
The exposure group included every patient who had engaged in a virtual visit for hypertension at any of Massachusetts General Hospital’s primary care clinics between December 12, 2012 and February 19, 2016 (n = 1051). We defined a “pre-visit” time period as 180 days before the in-person visit for hypertension. The “post-visit” time period was 180 days after the in-person visit for hypertension. The virtual visit occurred during this post-visit period.
The usual care group was drawn from all patients with hypertension who presented with principal diagnosis of essential hypertension (ICD9 401.9 or 401.1; shown to be a relatively accurate reflection of the true reason for a visit)9 to a primary care clinician’s office at any of Brigham and Women’s network of primary care clinics between December 12, 2012 and February 19, 2016 (n = 28,848). Because Brigham and Women’s primary care clinics do not yet have access to virtual visits but otherwise had similar information systems and are planning to implement virtual visits, they made for an uncontaminated usual care group. The pre-visit and post-visit time periods were defined the same as for the exposure group.
Propensity Score Matching Analysis
We used propensity score matching to create comparable cohorts of virtual visit and usual care patients. We used logistic regression modeling to create the propensity score of receiving a virtual visit (as if virtual visits were available at Brigham and Women’s network of primary care practices). We propensity score matched using age (continuous), gender (two classes), language preference (four classes), partner status (two classes), veteran status (two classes), smoking (two classes), race/ethnicity (four classes), health insurance coverage (four classes), total number of chronic conditions (continuous; of the 20 conditions considered chronic by the Health and Human Services Office of the Assistant Secretary of Health),10 visit year (continuous), mean pre-visit SBP (four classes), number of pre-visit antihypertensive medications (continuous), pre-visit primary care visits (continuous), pre-visit specialist visits (continuous), pre-visit emergency visits (continuous), and pre-visit inpatient admissions (continuous). We then used a SAS macro to perform nearest propensity score one-to-one matching with a caliper of 0.02.11 Unmatched characteristics are available in online eTable 1.
Because after propensity score matching on total number of chronic conditions the prevalence of diabetes was unbalanced in the groups, as a sensitivity analysis, we specifically added diabetes to the propensity score. This did not change any of our findings (online eTable 2).
For SBP, we calculated the mean mmHg of all SBPs obtained in the pre-visit and post-visit periods. SBPs included both in-person and virtual visits, where applicable. The change in SBP was pre minus post. We also examined in binary whether or not SBP decreased (“improved”) by at least 1 mmHg pre versus post.
For pre- versus post-utilization measures including primary care visits, specialist visits, emergency department visits, and inpatient admissions, we counted the number of each in-person event in the pre-visit and post-visit periods. We also examined in binary whether or not utilization changed by at least one event pre versus post.
Because many patients in the exposure group had well-controlled blood pressure, we performed a subgroup analysis that examined only patients without satisfactory blood pressure control (SBP > 140 mmHg).
We compared patients’ sociodemographic characteristics with Fisher exact tests for categorical variables and t tests for continuous variables.
We developed a multivariable linear regression model to examine the difference-in-differences in SBP change, primary care visits, specialist visits, emergency department visits, and inpatient admissions from the pre to the post period between virtual visit and usual care patients. We also assessed the count outcomes using negative binomial regression, but there were no substantive differences in the results, so we present the linear regression findings for ease of interpretability.
We fit additive (linear) probability models of change in SBP, primary care visit, specialist visit, emergency visit, and inpatient admission (the dependent variables).12,13,–14 A regression coefficient of an additive probability model can directly be interpreted as differences in the probability of technology use for a one-unit increase in the covariate corresponding to that coefficient. The additive models included an indicator variable for receipt of a virtual visit, an indicator for the pre and post period, and the interaction of receipt of a virtual visit with the pre and post period. The interaction term directly estimates the difference-in-differences, interpreted as the difference in outcome in the pre and post period between those with and without a virtual visit. We fit additive probability models for each outcome (five models).
We considered two-sided p values of less than 0.05 to be significant. We performed all analyses in SAS v9.4 (Cary, NC, USA). The Partners HealthCare Human Research Committee deemed this study exempt from review.