Most physiologic parameters, such as vital signs or activity, are routinely monitored a few times each day in hospital ward settings [1]. Some parameters, such as sleep, are not routinely monitored at all [2,3]. More frequent monitoring could allow for the timely identification of the deteriorating health of patients and spur efforts for improving patients’ overall health through increased sleep and activity. Since subtle changes in vital signs are often present 8 to 24 hours before a life-threatening event, such as intensive care unit admission or cardiac arrest, vital sign surveillance has the potential to detect clinical deterioration at an earlier phase, thereby permitting clinicians to make corrective interventions [4-7]. This includes identifying patients with poorly controlled pain and recognizing arrhythmias. The term posthospital syndrome has been used to denote the deleterious effects of acute illnesses that are compounded with poor sleep and low activity and occur during hospital stay [8]. Measuring sleep and activity could improve the recognition of such issues and encourage health providers to introduce interventions that improve patients’ experiences in hospitals by encouraging mobilization and to identify targets for sleep-promoting interventions [9-11]. In addition, access to other digital biomarkers (eg, heart rate, blood pressure, oxygen saturation, etc) would allow clinicians to determine underlying etiologies and make tailored interventions.

The rapid uptake of affordable wearables, such as fitness bands, may provide a method for continuously measuring sleep; activity; and vital signs, such as heart rate [12-15]. However, existing literature that describes wearable devices is mostly limited to ambulatory settings and focuses on the management of chronic diseases [16,17]. More inpatient data are needed on both the validity of wearables and patient adherence. Although wearable testing has been conducted with healthy volunteers, it will be important to validate these signals in inpatient settings, where algorithms for processing sensor data into digital signals, such as those for sleep, heart rate, and activity, may be less accurate [18]. Despite the proposed benefit of intensive monitoring, many wearable studies have found issues with patient adherence [18-20]. Adherence is a crucial barrier to acquiring data and can be influenced by device convenience, the comfort of use, and interaction requirements [19]. Studies of wearable devices worn by hospitalized inpatients have been limited by large dropout rates [20].

Although there have been systematic reviews of the monitoring of patients’ physical activity in hospitals [21-23], there are no reviews of the use of wearables that can reliably measure other health parameters. Therefore, in this review, we aimed to expand our search by including articles that used wearables to assess parameters other than physical activity and to assess the adherence of patients in inpatient settings.

For the purposes of this review, a wearable was considered to be any electronic device that has at least 1 sensor and can be worn on the body [24]. Wearables were examined for their ability to measure digital biomarkers, which are defined as digitally collected physiological and behavioral measures (eg, heart rate, average sleep duration, and daily step count) that explain, influence, or predict health-related outcomes [18]. Consistent with previous research, patient adherence was objectively assessed by reporting the mean proportion of patients who completed a given study [25]. The primary objectives of this review were to determine patients’ adherence to using wearable devices in hospitals and to examine the validity of wearable-derived biomarker readings.

A comprehensive search strategy was developed to identify articles on the three main concepts of our question—wearables, monitoring, and inpatients. The initial search strategy was developed for Ovid MEDLINE by using a combination of database-specific subject headings and text words (Multimedia Appendix 1). Additional key words were generated based on input from the subject specialists on the team, and the revised search strategy was customized for each database.

Searches of the following databases were executed on August 16, 2018: Ovid MEDLINE, Ovid MEDLINE Epub Ahead of Print and In-Process & Other Non-Indexed Citations, Cochrane Database of Systematic Reviews, Cochrane Central Register of Controlled Trials, Health Technology Assessment database (Ovid), and CINAHL with Full Text. The search in Ovid Embase was not executed until September 5, 2018, due to issues with the vendor’s August database reload. Additional search methods included reviewing the cited references of eligible studies via Web of Science (May 6, 2019) and the reference lists of eligible studies. There were no restrictions on publication period. Limits were imposed to ensure that only English-language studies and those with adult populations were included in this review. No other limits were applied to the literature search.

Records were screened by two reviewers (VP and RW) independently. For selected studies, full-text articles were obtained and evaluated for eligibility [26]. The eligibility criteria for inclusion in this review were as follows:

We excluded articles that were not considered original research, such as letters to the editor, comments, and reviews. We also excluded articles that monitored less than 3 patients, described the monitoring of a very specialized system in the body (eg, insole devices, ventricular assistive devices, and cochlear implants), involved the monitoring of patients in rehabilitation hospitals, or used wearables as tools for therapy (eg, insulin delivery).

Two reviewers (RW and VP) independently extracted the data and resolved any disagreements by discussing the findings and making a collective decision. The data extracted for each article included the year of publication, study setting and design, number of participants, gender ratio, mean age of participants, digital biomarkers measured in the study, average and maximum duration that the wearable was worn by participants in each study, and patient completion rate (the proportion of patients that wore the wearable for the minimum monitoring duration that was set by the study authors). For studies that used a reference standard, any participants who were missing data from the wearable or the standard were determined to be incomplete measurement pairs and were omitted from the final count of patients who completed the study. Furthermore, we extracted the types of wearables that were worn by the participants in each study along with the placement sites on the body. Devices were classified as medical grade (approved or cleared by the US Food and Drug Administration), research grade (typically used in research settings only), and consumer grade (used by general consumers).

Validation data were also collected for each article by assessing whether the authors compared the accuracy of their digital readings to a reference standard. To determine the validity of measures that were compared to a reference standard, correlation coefficients, mean differences, and limits of agreement were extracted from each study.

All articles that assessed for validated readings were independently assessed for their risk of bias by two independent reviewers (VP and RW) using a modification of the validation subscale from a checklist for assessing the methodological quality of studies on the measurement properties of health status measurement instruments (Consensus-Based Standards for the Selection of Health Status Measurement Instruments [COSMIN]) [27] (Table 1). All discrepancies were resolved by discussion and consensus. The quality evaluation included 5 study design and methodology components (the percentage of missing data, missing data management, adequate sample size, acceptable criterion comparison, and design or methodological flaws) and 1 analysis component (acceptable accuracy analyses). We rated the quality of each dimension as excellent, good, fair, or poor based on a priori modifications to the COSMIN validation subscale for scoring criteria that are appropriate for accuracy studies (Multimedia Appendix 2) [28].

Our literature search identified 2754 article citations. After excluding duplicate records, 2012 records were deemed eligible for screening. A total of 83 studies were selected based on abstracts and underwent full-text review. After applying our inclusion and exclusion criteria, 15 articles that described 14 studies were selected for this review (Figure 1).

All of the articles included were prospective cohort studies (Table 2) [4,11,13,14,20,29-38]. Overall, 9 different types of commercial wearables were described across the 14 studies, and 7 different health variables were assessed collectively by the 14 studies (Table 3). The wearable devices that were described by the studies came in various different forms and were attached to a range of sites on the body (Figure 2). A total of 13 articles included both men and women as the study participants; the other two papers assessed sleep changes in postpartum women [13,34]. Kroll et al [11,31] published two articles from the same study. Both articles analyzed different aspects of the continuous monitoring of inpatients (ie, they used the same cohort of patients) but were included as the same study entry in this review (Table 2).

Collectively, the mean patient completion rate across all 14 studies was over 90%. Of the 8 articles that included a qualitative analysis as a part of their methodology, 7 reported that wearables were well received by either or both patients and clinicians.

Of the 14 studies, 6 validated wearable measurements against another standard device or measure (Table 3). The studies conducted by Bloch et al [29], Gallo and Lee [13], and Steinhubl et al [32] used intermittent measurements (nurse or questionnaires) for their reference standard. Further, Breteler et al [30] used a continuous reference (continuous electrocardiography and impedance pneumography) to compare the wearable readings for heart rates and respiratory rates [30]. Weenk et al [4] and Kroll et al [11,31] validated their wearable readings against both intermittent and continuous reference measurements. Of the 9 wearables included in the studies, 6 were cleared or approved by the US Food and Drug Administration as medical devices (ViSi Mobile [Sotera Wireless], Hidalgo EQ02 [Equivital], wrist actigraphy [Ambulatory Monitoring Inc; Actigraph LLC], LifeTouch [Isansys Lifecare], Zephyr Biopatch [Medtronic], and HealthPatch [VitalConnect]).

Of the 6 studies in the risk of bias assessment, 4 were ranked as poor due to a small sample size (participants: N<30). The study conducted by Gallo and Lee [13] used sleep questionnaires as a reference measure and therefore received a fair rating for the “acceptability of reference” criterion, whereas the other five studies were ranked as excellent (ie, they used intermittent nurse readings or other validated methodologies). Further, in terms of assessing the accuracy analyses, only the study conducted by Bloch et al [29] did not report mean differences, correlations, and limits of agreement.

We conducted a systematic review that evaluated the utility of wearable technology in continuously monitoring hospitalized patients for a wide variety of health parameters. Our review focused on the breadth of devices used and the signals measured in hospitalized patients and included consumer, research, and medical-grade devices. There was evidence to support the use of Fitbit, ViSi Mobile, and the HealthPatch to measure heart rate [4,11,31], since the readings were validated against both intermittent and continuous reference standards. This review demonstrated that the validity of the data did not necessarily correlate with the classification of the device because even some medical-grade devices did not perform well and yielded data with wide limits of agreement. We found that only 6 studies validated the accuracy of wearable-derived health data from hospitalized patients by comparing the readings against a reference standard. Overall, the quality of most of these studies was excellent in terms of the reporting of missing data (6/6, 100%) and the use of acceptable accuracy evaluations (5/6, 83%). However, there was a considerable risk of bias in these studies due to the low number of participants in most of the studies (4/6, 67%). Many studies reported wide limits of agreement for other digital biomarkers, such as respiratory rate and blood pressure. Of note, we also found that the majority of studies (8/14, 57%) did not validate the studied device or parameter measured.

Of the various health parameters, the best evidence of validity was in the monitoring of heart rate in hospitalized patients. We also found that, in hospital settings, limits of agreement for medical-grade devices ranged from 16.4 to 21.8 BPM, whereas the limit for a Fitbit consumer device that uses photoplethysmography signals was 24 BPM. Further, during Fitbit-based continuous electrocardiogram monitoring, 73% of the readings were within 5 BPM of electrocardiogram readings. In a systematic review of 158 studies that measured heart rate by using consumer wearable devices, 71% and 51% of Apple Watch (Apple Inc) readings (used in 49 studies) and Fitbit readings (used in 71 studies), respectively, were within 3% of electrocardiogram readings in controlled settings [39]. Moreover, in 3 free-living studies, the wrist-worn Fitbit Charge had a mean absolute error percentage of 10% [39]. A systematic review of wrist-worn devices that measure heart rate via plethysmography found limits of agreement of 8.4 BPM at rest, 30.1 BPM while on a treadmill, and 41.5 BPM while cycling [40]. Overall, our findings found large limits of agreement for all devices, and inpatient results were consistent with the wide limits of agreement found in free-living environments or with activity.

We found that sleep only had a moderate correlation with sleep survey results from inpatient settings the use research and consumer devices. A recent systematic review of Fitbit-based sleep assessments found that readings from more recently developed devices correlated well with polysomnography readings for assessing sleep episodes [41]. It is unclear whether the lower correlation that we found was due to inpatient settings with high nighttime interruptions, patient factors that were perhaps associated with acute illness, or issues with sleep surveys (or a combination of these three factors) [3]. With respect to respiratory rate, 2 studies of 2 medical-grade devices provided limits of agreement. Wider limits of agreement were found in the study that had over 50,000 measurement pairs and used a gold standard (27 breaths per minute) compared to those in the study that had less than 100 measurement pairs and used clinician-reported vitals (13.4-19.3 breaths per minute) [30,32]. Additionally, previous studies found that medical-grade devices were only accurate under laboratory conditions or at-home conditions [42,43]. There was a limited number of studies on oxygen saturation, temperature, blood pressure, and fall risk.

There are a few limitations that should be noted for our systematic review. There is a considerable risk of bias, as the number of participants in the studies was low. Further, the studies included were observational in design and had a high degree of heterogeneity in terms of the objectives, populations, and outcomes reported. Thus, the data analysis methods were limited to broad categorization and the extraction of the common themes and trends that emerged from the results. Reports of wearable monitoring from individual studies should be viewed based on their methodological limitations. Although patient adherence has been found to correlate well with patients’ acceptability of wearables devices in inpatient settings, we realize that studying factors such as data loss, the duration of data gaps, and qualitative feedback from nurses and patients would further strengthen the generalizability of the results. Finally, it is important to note that wearable studies are being increasingly performed, and more relevant articles will become increasingly available.

This review also identifies gaps in knowledge that still exist within literature and provides information about what is required for further research. Specifically, the further validation of digital biomarkers by using gold standard comparators, such as polysomnography for assessing sleep and continuous electrocardiogram monitoring for assessing heart rate, is required. Ideally, large participant sample sizes and large numbers of measurement pairs within a population of interest should be used to assess parameters such as vital signs. The use of 2 reference standards to validate each health parameter, such as a heart rate, has also been recommended [44]. Moreover, data that are derived under real life conditions are still needed to better understand the factors that may contribute to between-patient heterogeneity when comparing the accuracy of wearable readings, such as those for patient activity, posture, gait type and velocity, locations of wearables, and patients’ diagnoses (eg, seizures). Future studies can aim to further qualify the process of retrieving data by using wearables to explore other barriers and avenues that might hinder the collection of reliable health information (ie, a weak Bluetooth connectivity, a lack of patient digital health literacy, the added burden that the process of taking wearable readings has on clinicians, the learning curve required to operate a wearable, etc) Finally, while we found that some digital biomarkers appeared to be valid for the monitoring of inpatients via wearables, we were unable to find any studies that supported the use of wearables in inpatient settings to improve clinical outcomes.

Overall, the assessment of studies in this review suggested that wearable devices show promise for monitoring the heart rate and sleep of patients in hospitals. The results show that many devices were not validated in inpatient settings, and the readings from most of the devices that were validated in such settings had wide limits of agreement. Further research is needed to determine the accuracy of the digital biomarker readings of hospitalized patients and to eventually determine whether wearable devices improve the health outcomes of hospitalized patients.