key: cord-0048687-6chebeqy
authors: Morgan, Anna U.; Balachandran, Mohan; Do, David; Lam, Doreen; Parambath, Andrew; Chaiyachati, Krisda H.; Bonalumi, Nancy M.; Day, Susan C.; Lee, Kathleen C.; Asch, David A.
title: Remote Monitoring of Patients with Covid-19: Design, implementation, and outcomes of the first 3,000 patients in COVID Watch
date: 2020-07-21
journal: NEJM Catal Innov Care Deliv
DOI: 10.1056/cat.20.0342
sha: 707368a2a94221f9aff8576bbc3140f0c56932a4
doc_id: 48687
cord_uid: 6chebeqy

An innovation team at the University of Pennsylvania Health System demonstrates how an automated text-messaging system can remotely monitor patients with confirmed or suspected Covid-19 at home and quickly support worsening patients with human care.

» Vigilant evaluation is required for these programs starting out -not just for new diseases like Covid-19, but because patient responses to clinical automation are not always anticipated.

» COVID Watch is a model that could be adapted to manage a variety of clinical conditions, such as hypertension, diabetes, or heart failure, where frequent human contact might be supplemented or partially replaced with automation.1

Although the most vivid images of the Covid-19 pandemic portray patients needing mechanical ventilation to survive, for every one patient requiring hospitalization, there might be 25 or more patients who are not sick enough to require hospitalization, but who could worsen quickly or be worried they might. Both groups might benefit from regular check-ins. In early March 2020, as the U.S. began to take seriously the threat of this disease, and as the clinical progression of the virus was still being understood, our team developed a strategy for watching over patients with confirmed or presumed Covid-19 at home.

We sought to solve several goals simultaneously:

1. Support a heterogeneous population of patients infected with Covid-19 who were remaining at home.

2. Identify and expedite the care of infected patients whose conditions were worsening.

3. Reduce health care personnel burden at a time when clinicians were being diverted to other tasks.

We developed a Short Message Service (SMS)-based automated remote monitoring program, called "COVID Watch," for Penn Medicine patients with presumed or confirmed Covid-19 infection. The texting program, which triggers calls from telemedicine clinicians when needed, was designed to supplement existing lines of care. We call our approach "automated hovering."2

The texting program, which triggers calls from telemedicine clinicians when needed, was designed to supplement existing lines of care. We call our approach "automated hovering."" " Figure 1 depicts the program's overall patient flow. Patients with presumed or confirmed Covid-19 are invited to enroll in COVID Watch. The program sends each patient automated twice-daily text messages to assess their health status. Based largely on dyspnea (shortness of breath), patients may be automatically escalated to a team of registered nurses available 24/7. Patients can also text "worse" at any time and be immediately escalated to a nurse, or end their participation by texting "bye." Nurses call escalating patients within one hour to assess and address their needs, and may refer them to the ED or an on-call team of physicians and advanced practice providers for additional telemedicine assessments. The program continues for 14 days from enrollment, at which point patients may opt for an additional seven days, after which they "graduate." Graduating patients are asked if they would recommend the program to others.

Patients are enrolled by a clinician entering the patient's cell phone number in the COVID Watch order embedded within our Epic electronic health record (EHR). Enrollment can take place at different times, including after a patient tests positive (by physicians staffing our results reporting system), during an ED visit, a call or telehealth visit with any UPHS clinician, or following discharge from an inpatient admission for Covid-19. We included patients with symptoms suggestive of Covid-19, such as cough and shortness of breath, because at the program's onset testing was limited, results were often delayed a week, and we felt the constellation of concordant symptoms was sufficient indication.

Enrolled patients received a text message inviting their participation and could accept or decline. Those declining were dropped. Those not responding were sent a reminder in the evening and were dropped after 48 hours of no response. Texts were sent in English at the start, and a Spanish version was added later. Figure 2 shows a sample SMS exchange. Patients who reported dyspnea in response to the twicedaily check-ins, or who texted "worse" at any time, triggered an EHR inbox message into a pool monitored 24/7 by registered nurses dedicated to COVID Watch. Nurses were expected to call patients within one hour.

Nurses were guided by a clinical algorithm that we developed to identify patients as (a) "emergent" and referred to the ED (e.g., patient cannot complete sentences), (b) "urgent" and referred to a 24/7 on-call team of physicians and advanced practice providers for a telemedicine appointment within 2-12 hours (e.g., breathing harder or faster than usual), or (c) "routine" and given immediate advice and asked to self-monitor or follow up with a PCP. Physicians supervised the clinical program 24 hours a day.

The COVID Watch team met seven days a week for the first eight weeks of the program, ramping down to three days a week. We built a comprehensive web-based dashboard, taking feeds from our EHR and the regional health information exchange, a database for sharing patient information on ED visits, admissions, and discharges. The dashboard enabled immediate review of every patient escalation and monitored patients enrolled in COVID Watch who visited any regional ED.

Patients who were sent to the ED but not admitted helped reveal how the automated program might be too sensitive. Patients who visited the ED without the program and were admitted revealed how it might not be sensitive enough. We paid careful attention to the latter patients, even though we saw those catches as successes, too; COVID Watch was never meant to be the only way patients could find their way to an ED but as a supplement to conventional approaches. Given delays in ascertaining out-of-hospital deaths, we could not identify false negatives -those patients monitored by the program who failed to escalate in our system and died at home.

We also conducted telephone interviews with patients who declined the program or dropped out before 14 days. These interviews, and our analysis of clinical escalations, helped us refine the algorithm, messaging, and clinical protocols. One such change broadened the focus on dyspnea to include fever, fatigue, nausea/vomiting, or chest pain. This alteration immediately resulted in a nearly 10-fold increase in escalation call volume, but did not result in any more patients referred to the emergency department. As a result, we reverted back to our dyspnea-focused algorithm after one day and excluded those escalations from the outcomes reported below.

The Institutional Review Board at the University of Pennsylvania reviewed the program and deemed it Quality Improvement.

Application-or web-based Covid-19 screening tools appeared early in the pandemic, relying first on travel history and later on symptoms alone. These programs recommended that high-risk individuals contact a doctor. An early hurdle was recognizing that we were trying to solve a few different problems. First, distinguishing who was infected and who was not was less important than distinguishing who was infected and needed hospitalization, and everyone else. Second, we did not want seriously ill patients to arrange their own care but wanted to connect them to care immediately. Third, as knowledge about Covid-19's course was evolving, we needed short-cycle quality assurance to oversee and modify the decisions of the algorithm and the clinical support that followed.

Our Center for Health Care Innovation (CHCI) conceived and led COVID Watch. Our first meeting to develop the program was on March 11, 2020; our first patient was invited on March 23; and by April 27 we had invited 3,000 patients. We were able to quickly build on Penn Medicine's previous investments in platforms and foundational experiences necessary for rapid innovation, including:

[1] Way to Health3 , 4 (W2H) is an automated platform created in 2010 with NIH funding to facilitate clinical trials of patient-engagement strategies and to become the platform for those strategies if successful. Way to Health combines messaging capabilities (SMS, e-mail, robocall) and other relevant automated features, such as electronic patient consent and secure data collection. Since its creation, W2H has been the foundation of over 170 clinical trials serving over 85,000 patients. It had already been integrated into our Epic EHR, so that individual clinicians could order W2H programs for their patients directly from the medical record.

[2] Penn Medicine had developed OnDemand,5 a telemedicine program designed originally to serve the urgent care needs of Penn Medicine employees, and then rapidly scaled to support the general patient community during the Covid-19 pandemic. This program provided the initial clinical infrastructure and experience for the team to feel comfortable offering automated engagement with infected patients at home.

[3] The innovation center had also developed Agent,6 a platform that merges data from our EHR and other digital sources. Agent applies data analysis and provides visualizations to support the monitoring of patient populations without needing to check individual patient records. Agent facilitated our ability to oversee this new program in real time.

The team comprised members from the Center for Health Care Innovation and the three programs listed above, and eventually included its own medical director, a nursing director, the clinical nurses (approximately seven FTE at peak), and several medical students. 

Each escalation was followed by a telephone call back from a nurse; 33 patients (6%) could not be reached. The mean and median response times were 16 and 9 minutes. Table 2 reports the clinical disposition of those responses: 83 patients (15.5%) prompted a recommendation to go to the ED and another 26 (5%) were already in the ED or admitted when the nurse responded. Emergency Department visit records could be found for 90 patients who had escalated within 12 hours of their ED visit, and of these 38 (42%) were admitted. An additional 107 patients were seen in the ED without escalating through the COVID Watch program, arriving there on their own or referred by another clinician. These patients had approximately the same rate of inpatient admission (39%-43%) as those who escalated through the program (42%).

As with an intention-to-treat analysis, we also followed patients who were enrolled in COVID Watch but declined or dropped out early from the program. These patients generated 58 ED visits; 52% of them were admitted. Table 3 illustrates these ED outcomes.

A SMS-based program in English and Spanish can't support patients who don't use text messages or don't speak those languages. Both of those problems are surmountable with more language options and the use of other digital communication tools, such as interactive voice recording. The latter is available through the Way to Health platform. The pressing need to quickly stand up a program for as many patients as possible, and to refine its design based on early experience, limited our ability to create a comparable control group. Nevertheless, robust existing data systems allowed us to evaluate the program's outcomes in real time. An initiative that uses centralized resources can seem like an additional cost to a system, but that is an accounting fiction: The system would need to care for these patients one way or another. The work required is determined more by the spread of the virus than by the organizational structure."

The automated component of COVID Watch used existing health system resources, including Way to Health and Agent. The follow-up call component was staffed by physicians and nurses whose normal clinical activities were displaced by the pandemic. As clinical staff return to their usual roles, we have faced the question of how to support continued operations. An initiative that uses centralized resources can seem like an additional cost to a system, but that is an accounting fiction: The system would need to care for these patients one way or another. The work required is determined more by the spread of the virus than by the organizational structure. We expect to continue centralized support for COVID Watch and related programs, like our Covid-19 resultsreporting system, because they provide economic efficiency and quality control.

COVID Watch combines automated SMS-based monitoring of patients with a 24/7 clinical back end. The program was designed to remotely observe patients with suspected or confirmed Covid-19 and to rapidly escalate to human care when needed. Our experience with the first 3,000 invited patients demonstrates efficiency and patient satisfaction. This is based on several factors, including automated check-ins that enabled seven FTE nurses to manage 1,000 sick patients 24/7 at the program's peak, and a high net promoter score.

As of June 21, 2020, COVID Watch has managed 4,249 COVID-19 patients at home. We also launched several companion programs, designed for overlapping purposes:

• Pregnancy Watch enrolled pregnant patients with Covid-19 and followed an essentially identical protocol, except that it escalated to experts in maternal fetal medicine. As of June 21, 2020, it had enrolled 138 patients.

• COVID Pulse was designed partly in response to reports that some patients with Covid-19 have little dyspnea, despite concerning hypoxemia (low blood oxygen).9 This program enrolls patients from emergency departments with depressed oxygen saturations, providing them with pulse oximeters, and escalates to care based on declines in measured oxygen saturation. As of June 21, 2020, it had enrolled 56 patients.

While we hope there will be no long-term need for COVID Watch, we believe that its broad uptake and support sets an example for future automated programs that facilitate patient care and communication while increasing efficiency."

While we hope there will be no long-term need for COVID Watch, we believe that its broad uptake and support sets an example for future automated programs that facilitate patient care and communication while increasing efficiency. We have a similar program running for patients with Chronic Obstructive Pulmonary Disease that also combines SMS-based monitoring with human back-up support, and we are in various stages of conceiving, developing, or evaluating programs for other conditions, including hypertension, congestive heart failure, and cirrhosis. 

Toward facilitated self-service in health care

Automated hovering in health care-watching over the 5000 hours

On the way to health

Subscribing to your patients-reimagining the future of electronic health records

The one number you need to grow

Net Promoter Score benchmarks

Silent hypoxia typically not the first symptom of COVID-19, other early symptoms should be monitored. American Lung Association

The authors would like to thank Catherine Armetta, P.J. Brennan