Discussion
Our study primarily demonstrates that individuals who responded to our survey tool were likely to be younger and healthier than those who did not respond. Although we noted an association between virtual engagement and improved patient outcomes, it is challenging to attribute this association to our intervention as opposed to baseline differences in responders, non-responders and those who refused enrolment. However, we do believe that our results support efforts to increase utilisation of telemedicine and automated self-care as they certainly demonstrate no worsening of adverse outcome risk.
Our survey-responding cohort was skewed towards younger people who may have had more access and familiarity with navigating technology. This is similar to associations found in other healthcare technologies, such as telemedicine video visits, in older adults.13 Women were more likely than men to respond to surveys but 15% less likely than men to experience the escalated care combined endpoint after adjustment for other confounders. This is not dissimilar to what other studies assessing survey response bias by gender have found, although no clear rationale for this discrepancy has been described to our knowledge.14 Responders also tended to be predominantly white, even though our overall cohort was primarily composed of black patients. As Black Americans are at greater risk of hospitalisation and death secondary to SARS-CoV-2 infection, elucidating reasons for decreased participation in virtual healthcare monitoring is critical.15 Patients who responded to our survey also had fewer comorbidities. We believe the reasons to this are twofold. First, patients with more comorbidities may be older and less familiar with technology, and second, patients with more comorbidities may be more likely to speak to specialist providers than to seek out care from a PCP.16
Our multivariable model showed that individuals who engaged in RPM or VOMC were about 50%–70% less likely to require hospitalisation or emergency care after controlling for EUA criteria, age, gender and race. Users of these services were skewed towards specific demographical groups who may have been healthier and at lower risk of hospitalisation compared with patients who were not enrolled or responders, and this effect size may be over-represented by our engaged cohort. To this point, when stratifying our combined endpoint, we found that ICU care and hospital admissions were decreased in the group using telemedicine and RPM surveys, but ED visits were not significantly decreased. For RPM survey–only responders, all three outcomes were significantly decreased. This suggests that these groups may have been overall healthier at baseline compared with those who did not virtually engage. Those who used VOMC visits only were not significantly likely to have a decrease in the combined or stratified endpoints, suggesting that these patients were likely sicker at baseline. Overall, fundamentally different baseline levels of health in each cohort may have skewed our statistical analysis. Although we did adjust for age, BMI and comorbidities, it is possible that the fact that our responding cohort was healthier at baseline had a disproportionate impact on our results and generated a large effect size statistically.
A less likely but possibly contributory explanation for this observed association between our responders and a reduction in ED visits could be a result of ‘auto-triage’ by RPM. Self-reported significantly abnormal vital signs or symptoms resulted in a triage call from a COVID-19 hotline nurse who would offer reassurance and self-management, a telemedicine visit and an ARC visit or direct the patient to the ED. In the absence of these services where less concerning vital signs or symptoms were filtered by multiple layers, patients may have self-referred to the ED. Our VOMC clinic is staffed by providers with experience in outpatient management of COVID-19, and our nurse triage team from our COVID-19 hotline has extensive experience in management of COVID-19 as well. Many symptoms that are expected from COVID-19 are alarming for patients, and separating out the most concerning signs or symptoms can be achieved by experienced practitioners. The specialised management from providers experienced with patients with COVID-19 or virtual engagement in general could have helped to reduce the need for some higher-level in-person services. However, we do acknowledge that given no medical interventions such as monoclonal antibodies were offered during the time of our study, we believe that there is insufficient biological plausibility to suggest that remote monitoring services reduced incidence of severe COVID-19 requiring ICU care. If our observation with a reduction in these services is in part reflective of virtual interventions, adaptation of virtual engagement methods may help reduce the burden of healthcare costs for patients who only require self-management and monitoring. Our virtual engagement strategies also allowed patients to receive specialised COVID-19 care from an academic medical centre even from remote locations when transportation may have been difficult due to geographical barriers or impossible given home isolation requirements from COVID-19 infection (online supplemental figure 1).
Multiple strategies to track patient symptoms in real time have been employed during the COVID-19 pandemic in the USA. For example, the Cleveland Clinic developed a phone app called ‘Care Companion’, which incorporated a similar workflow structure to our own RPM tool.17 The daily survey collects information on patient vital signs using a provided device and five symptom questions. This programme has been validated at other institutions, including Mass General Brigham, and is associated with decreased risk of admission to the ED or hospital.5 However, Care Companion was developed in collaboration with Epic, whereas our survey method has the potential to span multiple electronic medical record (EMR) systems. Additionally, our survey contained more questions about patient symptoms and did not require advanced mobile application familiarity. Mt. Sinai developed a similar application, called the Precision Recovery Program; however, this RPM programme was only offered to symptomatic patients when it was initially deployed.18 Their programme integrated both weekly telemedicine visits and daily RPM monitoring with escalation of care when worsening of symptoms was noted.
Our study has several notable strengths. First, we have a large patient population representing a wide range of patient demographics throughout several regions in Georgia. Our population also included patients who were tested at multiple environments—the emergency room, inpatient and outpatient—and was more likely to capture the large spectrum of COVID-19 presentations. Second, our survey tool did not require advanced technological knowledge, nor patient portal enrolment and was thus more accessible to our patients who might have been unfamiliar with technology or unwilling to download an application onto their phones. Third, our population included asymptomatic patients, whereas many RPM studies focused on actively symptomatic patients. This has important implications in characterising clinical sequelae for initially asymptomatic patients whose symptom severity worsens.
Several limitations to our study must be considered. First, because our survey was primarily delivered via text message, there is a possibility that patients who were non-responders thought they were spam and otherwise would have participated. We attempted to ameliorate this possibility by verbal communication about the survey format as well as clarifying language within the text message. Second, the 21-day implementation of the survey may have become repetitive for participants, particularly if they were asymptomatic or showed mild symptoms, and led to their discontinuation. We attempted to address this possibility by shortening the length of time surveys were sent to 14 days. Third, there is a possibility that the questions we posed were too numerous. We attempted to address this problem by asking patient history questions in the initial survey and narrowing down our focus to current symptoms in each follow-up survey with a particular focus on alarm symptoms that would be most predictive of patient outcome.
Additional biases in our analysis may be consequences of variable EMR maintenance differing by provider and facility. For example, our cohort is missing a portion of race data (1440 patients, 14.3% of overall cohort) due in part to EHR variability, and components of the Charlson Comorbidity Index (CCI) calculation may have been misclassified. Similar to most retrospective EHR-based studies, our outcome data also primarily reflect patient care provided at our hospital system and affiliates that we have EMR access to. We cannot conclusively say that non-responders or those who declined enrolment were more or less likely to visit an outside hospital; however, we suspect that the risk of misclassified outcomes applies to both groups equally. Finally, as a retrospective study, unmeasured confounding may exist due to initial indication of COVID-19 testing and evolving practices surrounding patients testing positive as more information about infectious sequelae, contact tracing and preventative strategies became available. Although our multivariate analysis controlled for age, race, gender and EUA criteria, we expect that additional confounding we were not able to control for contributed to the results. For example, technological proficiency and health literacy of the patients may have contributed to our observation of reduced need for higher level of care following diagnosis. Patients who are engaged virtually either through VOMC or Tonic surveys may have been more proactive about management of their health. Ultimately, a prospective, randomised trial would be required to address the possibility of unmeasured confounders and more definitely answer the question of whether virtual engagement affects clinical outcomes.
Future directions in our own survey implementation and for other health centres should focus on targeted outreach based on populations who are under-represented in responses, as well as those populations most impacted by COVID-19 and at risk of hospital or ICU admission. Our centre is planning to supplement our patient-based responses with wearable devices that may provide accurate, objective vital signs and oxygen levels to staff triage centres. Based on RPM programme responses and remotely recorded vital signs, we also hope to generate predictive machine learning algorithms to better refer patients to care centres before symptoms worsen. The strategies developed in this programme including a comprehensive registry have also facilitated the creation of predictive algorithms for identification of patients at highest risk of admission. Given that there are now evidence-based treatments to reduce risk of admission from monoclonal antibody infusion as we have also demonstrated at our own institution, we are poised to hopefully reduce the risk of admission of care through both technological monitoring and medical therapies.19 Additionally, with the increased understanding of implications of long COVID-19, survey questions may need to be revised to capture long-term consequences of COVID-19.
