We believe that monitoring a patient in their own environment during acute and transitional care can be less intrusive and more informative than in the hospital. However, we need to implement home monitoring collaboratively and iteratively, being mindful of the technology and the needs of the service, while managing the expectations of all involved [24]. The following recommendations are based on our experience.

It is important to select technology based on the program’s goals, rather than simply adopting what is available, previously purchased, or the norm on the hospital floor. Then, the technology should be actively studied on how it performs remotely. Regulatory clearance (eg, Conformité Européenne [CE] marking or Food and Drug Administration 510(k) clearance) is the minimum bar for any clinical device, but the laboratory validation studies required for clearance give little indication of how well it will perform in the home. Pragmatic testing in the context in which the technology will be used is essential to ensure it will be fit for purpose. This includes ensuring equitable provision for patients with low levels of digital literacy or access [25]. Challenges such as the provision of infrastructure or subsidization of access are more easily addressed on a governmental level. However, technology providers can work to design accessible standards, develop specific training programs, incorporate translation services into platform communications, and ensure that cultural sensitivities are respected [26]. Building in free-roaming cellular connectivity to devices, rather than relying solely on patients’ own internet connections, is an effective way to allay concerns about incurred costs. In all cases, user experience testing is key to validating assumptions about the target population. For example, one might have assumed that older patients would struggle with remote monitoring technology. In a comparison between 316 patients aged over 75 years and 541 younger patients using remote monitoring equipment, the older group did indeed rate themselves as more likely to avoid technology or delay adoption (82% vs 56%) and scored the technology as marginally less easy to use (ease-of-use score 5.5 vs 6.2) [27]. However, the older patients exhibited higher adherence to wearable use (95.3% vs 93.3%; P<.001) and equivalent adherence to remote monitoring tasks such as blood pressure measurement. The obvious implication was that patients should not be excluded because of age, but the more nuanced finding was that patient education should be focused on building confidence in the technology as much as skill acquisition.

Once a candidate technology is chosen, how it will work in the broader ecosystem should be considered, including interoperability, cybersecurity, and the flow of aggregated data into the EHR. Specifically, this may include procedures to annotate or amend data flowing from the home into the EHR if the data are missing or erroneous. This is not always straightforward. As Figure 1 illustrated, there is a compromise between accuracy and acceptability: no wearable device will ever be completely accurate all the time and may occasionally transmit incorrect or artifactual data to the medical record. This is also true of manual measures in the hospital, which have considerable interobserver variability, but these are more of a “cultural norm,” and in the home, there are usually no additional measures to corroborate the data from the wearable [19]. One approach is to annotate remotely collected vital signs with a measure of confidence that the vital signs accurately reflect the patient’s physiology, based on the strength and regularity of the waveform data, the degree of interference (eg, from movement), and the percentage of data available in each window of time.

Before implementing remote monitoring, it is important to educate hospital clinicians to expect different and more variable data. Gaps in continuous data from the home are normal. Perfect accuracy or coverage is not expected (or necessary) to make good decisions, even in acute care. Our nursing team has been able to safely monitor 5515 patients at home with a median (IQR) adherence to continuous monitoring devices of 79% (77%-81%; defined as the percentage time data transmitted over length of stay)—in other words, with approximately 20% of data “missing.” This remains the case even for the most vulnerable populations: in a cohort of 40 patients with hematological malignancies receiving chimeric antigen receptor T-cell (CAR-T) therapy managed acutely at home (median length of stay 14.7, IQR 10.4-27.6 days), overall adherence to continuous monitoring devices was 79% (IQR 68%-88%) [28]. Median (IQR) survey completion was 50% (7.7%-80%) and blood pressure adherence was 95% (69.8%-100%). A total of 25 patients developed cytokine release syndrome while at home but were safely managed, despite incomplete data. We advise repeated training sessions as clinicians gain familiarity with this new stream of data. Patients and caregivers should be trained to use the equipment properly, accounting for their knowledge, motivation, skills, systems, and behaviors [13]. If the collected vital signs are visible to patients and carers, then a sufficient degree of education and annotation should be insured so they can react constructively to their data without undue concern [25].

It is recommended to move away from the hospital practice of applying “standard monitoring.” Instead, clinicians should be encouraged to consider home monitoring as a diagnostic intervention with downstream clinical, operational, and financial consequences. An appropriate “dose” of monitoring should be prescribed to capture meaningful change. A monitoring prescription might be used for intermittent or continuous data capture, specifying a bespoke cadence of data collection, adherence, duration, alarm thresholds, escalation, and discharge criteria. The prescription may need to be adjusted over time to reflect the patient’s clinical trajectory. As home monitoring matures, we can evolve dose-response curves to titrate monitoring across different use cases. This dovetails with the previous point regarding education and setting expectations. In the outpatient CAR-T program described above, 30 patients triggered a total of 670 clinical alarms (0.5 per patient-day) over 4 months of monitoring [28]. These alarms prompted 1102 calls by monitoring nurses to patients, 63% of which were during nonclinic hours. The alarms were set to trigger when the median of a particular vital sign, aggregated over 30 or 60 minutes, breached a set threshold. When a 15-minute aggregation window was tried in an analogous program, 6 patients triggered a total of 436 alarms in 1 month without a notable difference in outcomes to the first program.

It is beneficial to individualize alarms in the home. In the hospital, a patient’s baseline vital signs, especially their variability, will be suppressed as they rest in bed. In the home, there will be more variation. We advise using alarm thresholds and trends tailored to individual patients and based on their clinical condition, risk, and social circumstances, rather than generic hospital “normal” values [29]. Their baseline values should be established early in their home stay, and then the focus should be on deviations from that baseline, rather than absolute values [30]. This recommendation is perhaps best illustrated with an inpatient example, where a “gold standard” comparator (standard nursing care) can be used. In a study of 28 patients receiving CAR-T therapy in the hospital with a remote monitoring device worn as a “black box,” a simple alarm set at baseline temperature plus 2 SDs allowed the detection of temperature disturbances indicative of cytokine release syndrome a median of 184 (IQR 53-312) minutes earlier than standard care [31]. By comparison, the use of a fixed temperature threshold only improved detection by a median of 22 (IQR –43 to 113.5) minutes without a significant increase in specificity (Figure 4 [32]). Using trend-based measures is also an approach to deal with the variability in the peripheral devices described in Figure 3. While Figure 3 concerned blood pressure, in a systematic review and meta-analysis of 75 studies, Niven et al [33] demonstrated strikingly broad limits of agreement between peripheral and central measures of temperature, especially among adults with fever (–1.44 to 1.46 °C). In other words, peripheral measures of temperature in 2 patients with the same core temperature could vary by ~3 °C. This supports moving to a trend-based diagnosis from the fixed hospital threshold of 38 °C in febrile illness.

In the hospital, we can use alarms that are purely sensitive (high risk of false positives and low risk of false negatives). Alarm fatigue is an issue, but it is straightforward to examine the patient following an alarm [34]. In contrast, a false alarm in the home may entail waking the patient or caregiver or dispatching a paramedic unnecessarily. We must therefore introduce more specificity. However, if we push either sensitivity or specificity too far, there will be a loss of trust from the patients and clinicians, financial penalties, and degradation of clinical care, as in the oncology example above (Figure 5).

We must find a balance of sensitivity and specificity, while being mindful of our clinical risk tolerances and our capacity to respond to alarms. All decisions should be based on new and variable data. Accepting this balance allows us to give patients space to recover in their own home and grants us a rich stream of diagnostic information. It is the crux of the mindset shift to home monitoring. It also requires establishing a feedback loop to differentiate useful and nonuseful alarms, adjusting the balance of sensitivity and specificity accordingly. This titration can be challenging, as there may be a gap between an alarm sounding and its utility being understood. Furthermore, those monitoring the data, assessing the patient, and modifying alarms might all be different individuals in different locations. In all cases, escalation criteria should be agreed in advance, being mindful that clinicians receiving the referral may not be intimately familiar with the nuances we have described above. We suggest erring toward sensitive settings at the start and increasing specificity as confidence grows.

It is important to not rely solely on alarms. Instead, teams should conduct “rounds” on a virtual care patient’s data as often as their condition warrants to become familiar with the patterns in their vital signs. Our monitoring team conducts a “round” on patients every 4-6 hours, as they might on the hospital floor. A “round” involves reviewing the patient’s vital signs and any asynchronous communications. This helps nurses establish patterns that differentiate the routine from the pathological, as well as picking up issues like incorrect device wear. Nurses also use surrogates for examining a patient. For example, compensatory changes in other vital signs, measures of motion, postural change, speed of vital sign recovery following activities of daily living, and adherence to monitoring all paint a picture of a patient in the home. These observations can then be fed back to patients’ care teams during multidisciplinary “huddles” [35]. Our nurses have anecdotally reported that hypoxia and rising temperatures are most frequently spotted early by rounding, particularly when the patient’s health care providers have chosen alarms that maximize specificity. As one nurse described:

Above all, we must recognize the learning curve and acknowledge the initial uncertainty and stress for care teams with this new stream of data. Clinicians should be monitored as they develop meaningful relationships with their RPM patients [36]. With time, they will develop confidence to not disturb patients, such as the patient referred to in Figure 2, with every fluctuation in vital signs—even in the acute setting. While desirable, this is also medicolegally vulnerable position, as the clinicians would opt not to respond to values traditionally considered “abnormal” in hospital, to give patients the space to recover. Malpractice is judged as deviation from an accepted standard of care. While there are now national figures for escalation and mortality rates (6.2% and 0.5%, respectively, in the United States), there are no standards of care for responding to data from the home [2,37]. If we default to hospital standards of care (rapid response to every change in vital signs and alerting senior staff), we either undermine care at home by overcontacting patients or risk a monitoring care team being in legal jeopardy for appearing to be unduly lax. So, we must set new standards for care at home, supporting clinicians in their decisions, building facilities for them to easily document their judgment calls, and constructing achievable escalation pathways around them. Standards and governance need to develop alongside a program as it evolves, with experienced clinicians mentoring those joining the program. Given all the limitations and subtleties we have described, we advise caution in entrusting too much triage, too early, to machine learning alone.

There is often insufficient clarity around data ownership in an RPM program. There is a need to establish early on what data belong to the patient, the technology provider, and the health system [25]. This should be considered for both identifiable and deidentified data. This is particularly pertinent in the context of program evaluation and health care research. The most useful data are typically those labeled with symptoms or behavior by the patient, with quality and completeness by the technology provider, and with clinical context by the health system. If all 3 cannot be physically and legally united, then evaluators and researchers would not hold all pieces of the puzzle.