Use telehealth as needed: telehealth substitutes in-person primary care and associates with the changes in unplanned events and follow-up visits

This study adopted a retrospective design and used electronic health records and claims data for Medicare fee-for-service beneficiaries served by an Accountable Care Organization [44] (ACO) in the Midwest of the United States. The study period was defined as April to December 2020, and the baseline period was from March 2019 to February 2020. March 2020 was excluded since the COVID-19 pandemic started in March and made the month less comparable to other months in the study time window. Beneficiaries who were 65 years or older and with a primary care physician (PCP) in the ACO provider list at the beginning of the study period were included in the analysis. Telehealth was coded from date of patient receipt of primary care services by audio or video telecommunication using standard billing (CPT/HCPCS) codes [4, 17, 26, 45‐47].

Two types of the data were used for analysis. First, the non-COVID related care utilization data was collapsed into patient-month level and was used as longitudinal data throughout the observational period: 12-month baseline and 9-month study period. Second, the outpatient encounter-level data was organized by beneficiary and encounter date. One encounter represents one primary care visit (in-person or telehealth). A patient could have none, one or more than one in-person or telehealth encounter in any month. This encounter-level data was used to compare telehealth and in-person visit utilization and follow-up patterns. This study is exempted by the Institutional Review Board in the University of Wisconsin-Madison. Reporting of the study follows the STROBE checklist for observational studies.

For the patient level analysis, outcomes are the occurrence of unplanned events (ED visits and hospitalizations) and the total unplanned Medicare payment ($) in each month. To differentiate care utilization at the intensive and extensive margin, the occurrence of unplanned events was measured as the number of ED visits or hospitalizations as well as whether there were at least one ED visits or hospitalizations in a month. For the encounter-level analysis, outcomes are the number of days until the next encounter and further, whether the next encounter happened within the 3-, 7-, 14- and 30-day frame.

In the patient-level analysis, the two major explanatory variables were a binary indicator of the study period relative to baseline, and a categorical variable to differentiate patients by their outpatient utilization throughout the entire study period: whether a patient had no visit, telehealth visits only, in-person visits only, and visits of both types—in-person and telehealth. Different from other existing studies which used telehealth utilization status as the sole indicator [24‐27], this study considered the co-existence of in-person visits and categorized beneficiaries into four mutually exclusive subgroups. We hypothesized that patients who used only telehealth and those who used both types of services were different and should not be grouped together when compared with the patients who did not use telehealth (the in-person only group). For the encounter-level analysis, the two major explanatory variables were a binary indicator of the study period during the pandemic relative to the pre-pandemic baseline period, and an indicator of the encounter type: telehealth relative to in-person visits.

For both patient- and encounter-level analysis, control variables included sociodemographic variables, health conditions, care utilization at the baseline as well as month fixed effects to control for seasonal trends. Inclusion of these variables are based on data availability and the Andersen's Model of Health Care Utilization [48] including predisposing factors such as a beneficiary’s age group (65–74 years old, 75–85, and 86 +), sex (0/1), race/ethnicity (non-Hispanic White or minorities); socioeconomic factors such as Medicaid enrollment (0/1), disability entitlement (0/1) and the geographical residence category (urban, suburban, large town, small town/isolated rural); and need characteristics such as the hierarchical condition categories (HCC) score, an indicator of having three or more chronic conditions (0/1), the total number of unplanned event counts (ED visits and hospitalizations) and payments at the baseline period.

For the patient-level analysis, logistic regression models were used to predict the probability of having at least one ED visit or hospitalization per patient per month. Negative binomial models were used to predict the number of ED visits or hospitalizations. Ordinary least square models were used to predict the average logarithm of unplanned payments. All models were controlled for patient characteristics at baseline and the month fixed effects. The predicted probabilities, number of unplanned events and dollar value of payments were reported.

For the encounter-level analysis, we first plotted the distributions of the number of days until the next visit for all encounters by time (pre-pandemic period vs. during) and the encounter type (telehealth vs. in-person visits). It was hypothesized that if a patient used either telehealth or in-person encounter type as a substitute for the other, the distributions of time until the next visit should be comparable between the two encounter types. In addition, we used Poisson regression models to estimate the time duration until the next visit and used logistic regression models to estimate the probabilities that the next visit happened within 3, 7, 14 and 30 days since the last visit. These models were adjusted for the pandemic time indicator, encounter type and month fixed effects. The incidence rate ratios (IRR)/odds ratios (OR) from the regressions, the predicted number of days until the next visit and the probabilities of 3-, 7-, 14-, and 30-day follow-up were reported.

Including patient-level characteristics at the baseline period and the month fixed effects in all analysis helped to absorb the potential confounding bias at the patient-level or due to seasonal trends. Robust standard error was also used to refine estimates. Records with missing data, which accounted for less than 1% of the study sample, were dropped from the analysis.

The study sample included 338,872 patient-month records and 134,375 outpatient encounters during the observational period (March 2019 to December 2020) for a total of 16,222 beneficiaries. The study sample had a mean age of 76 (standard deviation SD 6.5) years old, 61% female, 96% non-minority White, 6.5% with Medicare & Medicaid dual coverages, 4.3% with disability entitlements, 70% from urban areas, 19% suburban and 11% from large, small town or remote rural areas (Table 1).

During the 9-month study period (April to December 2020), 18.3% (N = 2,976 out of 16,222) of the beneficiaries had no outpatient visit, 18.1% (N = 2,929) had only telehealth visits, 29.9% (N = 4,853) had only in-person visits, and 33.7% (N = 5,464) had both types. Comparing across the four outpatient encounter utilization categories by baseline characteristics, those beneficiaries who used both telehealth and in-person visits were older (76 vs. 75 yrs.), had a higher average HCC score (1.43 vs. 0.93–1.02) and were more likely to have 3 or more chronic conditions (64% vs. 31–48%) than those from the other three categories. Comparing between the telehealth only and in-person only groups, the in-person only group had more male beneficiaries (42% vs. 38%), more between 75–85 years old (34% vs. 32%), less 86 years or older (8.8% vs. 10.4%) or from urban areas (65% vs. 74%). The average HCC scores were comparable but the probability of having 3 or more chronic conditions were lower in the in-person only group than the telehealth only group (41% vs. 48%). These results provided initial evidence that less healthy beneficiaries or those with more clinical needs used both telehealth and in-person visits, while those healthier ones chose either type to use by their preferences (Table 1).

After risk adjustment, the average predicted probability of having at least one ED visit per month after adjusting all patient characteristics during the baseline period (March 2019 – February 2020) were comparable across the three healthier groups (2.02%, 2.21% and 2.32% for the no-visit, telehealth-only and in-person only group). The group that used both types of visits had a significantly higher baseline probability at 3.5% (95% CI [3.4%, 3.6%]). In comparison, during the study period, the predicted probability for no visit, telehealth-only and the in-person only groups significantly decreased relative to their baseline levels (0.77% CI [0.7%. 0.9%], 1.23% [1.1%, 1.4%], and 2.01% [1.9%, 2.2%]). But for the group that used both types of visits, the predicted probability of have at least one ED visit per month did not significantly change from the baseline (3.52% [3.4%, 3.7%]) (Table 2 & Supplementary Table 1).

The predicted number of ED visits before and during the pandemic across the four care utilization groups followed similar patterns. In the baseline period before the pandemic, the estimated average number of ED visits per 1,000 patients per month were comparable at 22.5, 24.6 and 26.1 (±1.8) visits for the no-visit, telehealth-only and in-person only groups and was significantly higher at 40.1 (95% CI [38.4, 41.7]) visits for the both-types group. During the pandemic, the first three groups significantly reduced the number of ED visits (8.2, 13.2 and 21.9, ±1.6) while the fourth group did not show significant change from before (41.8 CI [39.6, 43.9]) (Table 2).

Comparing across the no-visit, telehealth-only and in-person only groups before and during the pandemic, the three comparable groups in ED visits during the baseline became significantly different from each other during the pandemic. The in-person group experienced the least reduction in ED visits, yielding the highest probability and number of ED visits among the three, while the no-visit group was the opposite. Comparing between telehealth-only and in-person only groups, the two groups had comparable baseline health conditions and ED visits, but the in-person only group had significantly higher predicted probability and number of ED visits than the telehealth-only group during the pandemic. Overall, these results suggested that patients used either telehealth or in-person visits or both depending on their clinical needs and availability and treated the two types of care delivery modes as a substitute to each other.

The changing patterns of the predicted probability and number of hospitalizations across the four care utilization groups were similar to that of the ED visits in the baseline period but were different during the pandemic. In the baseline period, the number of hospitalizations for the no-visit, telehealth-only and in-person only groups were comparable among each other (at 11.0, 12,7 and 12.9±1.6 hospitalizations per 1,000 patients per month respectively) and were significantly lower than the both-types group (17.8 CI [16.8, 18.9]). During the pandemic, the no-visit and telehealth-only groups had significantly fewer hospitalizations relative to the pre-pandemic period (5.5 and 8.1±1.6). The in-person only group showed no significant changes (12.3 CI [10.9, 13.7]) and the both-types groups had a significantly higher number of hospitalizations than before (23.0 CI [21.4, 24.6]) (Table 2).

In the baseline period, the average monthly payment per person for the telehealth and in-person groups were comparable between each other ($7,326 and $7,184) but were significantly higher than the no-visit group ($4,772) and lower than the both-types group ($13,449). During the pandemic, the no-visit, telehealth only, and in-person only groups had lower payments than pre-pandemic levels ($1,896, $3,778, and $4,838 respectively), but the both-types group had no significant changes from before ($12,588). Further, the estimated payments for the telehealth only and in-person only groups became significantly different from each other during the pandemic (mean difference $1,060, CI [$957, $1,162]) (Table 2).

The overall distribution of the number of days until the next visit among all encounters were comparable before and during the pandemic and between in-person and telehealth visits. More than 60% of the next visits happened within 30 days. Proportionally more next visits happened on the 7^th, 14^th, 21^st and 28^th day from the initial visit (Fig. 1).

After adjusting for patient characteristics and month fixed effects, both telehealth and in-person visits had fewer number of days until the next visit during the pandemic relative to in-person visits before the pandemic (33.4 and 31.2 days respectively vs. 43.9 days; OR 0.76 [0.75,0.78] and 0.71 [0.70, 0.72]), but the two visit types were not significantly different from each other in number of days until the next visit. Similarly, the predicted probabilities of 3-, 7-, 14- and 30-day follow-up were higher during the pandemic period than before (OR 1.36–1.74). Comparing between telehealth and in-person visits during the pandemic, telehealth visits had comparable 3- and 7-day follow-up rates (9.2% [8.6%, 9.7%] vs. 9.3 [8.9%, 9.7%] for 3-day; 21.8% [21.0%, 22.5%] vs. 23.5% [22.0%, 24.1%] for 7-day) but lower 14- and 30-day follow-up rates than in-person visits during the pandemic (37.8% [36.9%, 38.6%] vs. 40.6% [40.1%, 41.3%] for 14-day; 62.4 (61.6, 63.2) vs. 65.1% [64.5%, 65.7%] for 30-day) (Table 3).

This study has a few limitations. First, the study sample came from a single care system and could not represent the general aging population nor Medicare beneficiaries from managed care plans. However, focusing on one care system allowed us to leverage richer information from both electronic health records and claims data for analysis [49, 50]. Second, findings from the study suggested that telehealth was used as a substitute for in-person services and did not lead to sooner or more frequent services in the future. But as with many other observational studies, we cannot establish causal relationship between telehealth and care outcomes, such as unplanned events and follow-up visits. Some study designs to construct comparison groups or with randomized interventions would help. Third, for the analysis on the time duration until the next visit, we can only group the outpatient encounters by beneficiary and encounter date and cannot guarantee that any two consecutive encounters were for the same healthcare concerns. Dose response relationship across follow-ups beyond the immediate next visit was not considered either. Similarly, the data did not include information on the waiting time from requesting an appointment to being seen for telehealth or in-person visits, nor were data available to discern whether an ED visit or hospitalization was just a one-off event or due to a patient’s bypass of primary care. Using additional information on primary diagnosis and reason for the visit will help refine the results. Lastly, results from this study were based on the actual utilization data. Patients who faced barriers to access telehealth, such as technology or health literacy, had to stay with in-person services or no service at all during the pandemic period. As such, results from the study were conditional on telehealth availability and accessibility. More system and community efforts are expected to ensure equitable access to telehealth as an alternative to traditional in-person services in the long run.