Effect of Telemedicine on Cardiology Care Patterns

Since the COVID-19 pandemic hit the US, healthcare providers have turned to telemedicine visits (by video and telephone) as a valuable alternative to standard in-person visits. Now, more than two years later, some providers returned to mostly in-person care, while others kept using telemedicine as their primary modality. A major question facing Stanford Healthcare’s Cardiology clinics as well as medical clinics around the country is whether and to what degree telemedicine should be used, particularly for new patient visits. 

To help address this question, our SURF project aimed to develop a methodology for estimating differences in patient outcomes and patterns of care for patients who had their initial visit via telemedicine versus in person. We focused exclusively on patients who had a new patient visit at a Stanford clinic after June 2020 with one of the following subspecialties: General Cardiology, Electrophysiology, Interventional Cardiology, Heart Failure, and a primary visit diagnosis in one of the following groups: atrial fibrillation / flutter, chest pain, coronary artery disease, dyslipidemia, heart failure, hypertension, palpitations, pre-op. The main outcome of interest we studied was how many follow-up visits (via telemedicine or in person) the patient had in the six months following their initial visit.

Some preliminary analysis found that patients who had their first visit via telemedicine were slightly more likely to have private insurance than those who came in person. Additionally, we found that providers within the same subspecialty varied significantly in how much they used telemedicine (0 – 100%), even when seeing patients with the same primary visit diagnosis. Prior interviews with clinic schedulers (from a past SURF project) indicated that providers are allowed to dictate what share of their new patient appointment slots were for telemedicine versus in person. Some other important facts from these interviews were that patients were assigned the provider who had the next available appointment (of either modality), then given an option to choose the alternative modality at a later date with the same provider, if the provider indeed used both modalities. 

This led us to design an instrumental variables approach to estimating the effect of telemedicine on six-month follow-up visits. While we can’t control for a patient’s internal preference for one modality or the other, we can assume that they were randomly assigned to a provider whose availability of appointment slots would influence the modality the patient ended up receiving. In other words, the assigned provider’s fraction of telemedicine visits is the instrument. We decided to apply this instrument through the traditional two-stage least-squares regression method, as well as a newer machine learning-based method called instrumental forest.

The two-stage least-squares method estimated that having telemedicine as the modality for the initial patient visit had a small but statistically significant effect of -0.13 on the number of follow-up visits (95% confidence interval = [-0.25, -0.01]. The instrumental forest method, however, found no statistically significant difference between the modalities, with an effect size of -0.05, and a 95% confidence interval of [-0.13, 0.03].

Moving forward, we plan to 1) continue to refine our methods, 2) test the robustness of our results on different subsets of the data and with different controls, 3) test for heterogeneous treatment effects, and 4) test our methodology on different outcomes of interest, such as prescription fill rates and subsequent hospitalizations.