What Is the Impact of Ambulatory Vital Sign Monitoring on Deterioration Detection and Related Clinical Outcomes in Hospitalised Patients: a Systematic Review and Meta-analysis.

Background: Timely recognition of the deteriorating inpatient remains challenging. Ambulatory monitoring systems (AMS) may augment current monitoring practices. However, there are many challenges to implementation in the hospital environment, and evidence describing the clinical impact of AMS on deterioration detection and patient outcome remains unclear. Objective: To assess the impact of vital signs monitoring on detection of deterioration and related clinical outcomes in hospitalised patients using ambulatory monitoring systems, in comparison with standard care. Methods: A systematic search was conducted in August 2020 using MEDLINE, Embase, CINAHL, Cochrane Database of Systematic Reviews, CENTRAL and Health Technology Assessment databases, as well as grey literature. Studies comparing the use of AMS against standard care for deterioration detection and related clinical outcomes in hospitalised patients were included. Deterioration related outcomes (primary) included unplanned intensive care admissions, rapid response team or cardiac arrest activation, total and major complications rate. Other clinical outcomes (secondary) included in-hospital mortality and hospital length of stay. Exploratory outcomes included alerting system parameters and clinical trial registry information. Results: Of 8706 citations, 10 studies with different designs met the inclusion criteria, of which 7 were included in the meta-analyses. Overall study quality was moderate. The meta-analysis indicated that the AMS, when compared with standard care, was associated with a reduction in intensive care transfers (risk ratio, RR, 0.87; 95% condence interval, CI, 0.66 to 1.15), rapid response or cardiac arrest team activation (RR, 0.84; 95% CI 0.69 to 1.01), total (RR, 0.77; 95% CI 0.44 to 1.32) and major (RR, 0.55; 95% CI 0.24 to 1.30) complications prevalence. There was also association with reduced mortality (RR, 0.48; 95% CI 0.18 to 1.29) and hospital length of stay (mean difference, MD, -0.09; 95% CI -0.43 to 0.44). However, none were statistically signicant. Conclusion: This systematic review indicates that implementation of AMS may have a positive impact on early deterioration detection and associated clinical outcomes, but differing design/quality of available studies and diversity of outcomes measures limits a denite conclusion. Our narrative ndings suggested that alarms should be adjusted to minimise false alerts and promote rapid clinical action in response to deterioration.


Background
In the United Kingdom, the use of physiological early warning scoring (EWS) systems (which measure "standard" vital signs such as pulse rate, respiratory rate, blood pressure, oxygen saturation, and temperature) is still common practice in general wards, together with a graded response such as referral for a senior review or increasing monitoring frequency [1]. This frequency of observations is generally guided by the clinical condition of the patient, and due to the requirement of manual physiological measurements, it can be time-consuming for healthcare professionals [2]. As a result, the optimal monitoring frequency is often not achieved [3], limiting the e cacy of intermittent monitoring systems dependent on the frequency of manual observations [4]. Furthermore, even when the ideal frequency is achieved, patients can deteriorate between observation sets [5]. Higher risk patients are often continuously monitored (for example in critical care), improving early detection of deterioration [2]. However, in the UK, continuous monitoring is not commonly used in the ward environment [6], although one study suggests it may be feasible and cost-effective in surgical wards [7], with the potential to improve patient outcomes when compared to intermittent monitoring [5].
Despite the potential to promote earlier detection of deterioration, limitations in continuous vital sign monitoring technology can pose a barrier to implementation [2], such as restriction of patient mobility and independence due to wires and static devices [6,8]. In response to this need, commercially available wearable monitoring devices are evolving rapidly [9]. Wearable devices may provide an alternative to static wired continuous monitors and offer a bridge between bedside wired monitoring and intermittent manual measurements. This development has the potential to promote patients' mobility and comfort while reducing nursing time and improving the early detection of abnormal physiological parameters [10].
A recent meta-analysis assessed the impact of multi-parameter continuous non-invasive monitoring in hospital wards, including wired static devices, suggesting a 39% decreased mortality risk in monitored patients compared to those receiving standard care (intermittent manual observations), it also suggested a trend of reduced intensive care unit (ICU) transfer, rapid response team (RRT) activation and hospital length of stay (LoS) [11]. The validation, feasibility, costs and clinical outcomes of 13 different wearable devices were assessed in another systematic review, which demonstrated that the majority of studies were still at the validation and feasibility phases [12], emphasising the lack of evidence assessing the impact on economic and clinical outcomes as there is still uncertainty around the impact of ambulatory monitoring systems (AMS) in the hospital environment, hindering its implementation and clinical use [13]. Our review focused on these ambulatory monitoring devices and/or systems implemented inside the hospital (inclusive of all specialities, acuity and ages).

Objective
The objective of this systematic review and meta-analysis was to assess the impact of vital sign monitoring on the detection of physiological deterioration and related clinical outcomes of hospitalised patients using ambulatory monitoring systems in comparison with standard care.

Methods
This systematic review was registered with the International Prospective Register of Systematic Reviews (PROSPERO) on the 10th July 2020, registration number CRD42020188633 [14]. This review was reported following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) checklist (Appendix 1) [15]. The full systematic review protocol was published prospectively [16].

Primary Outcomes
This study aimed to compare the impact of ambulatory monitoring systems on deterioration detection and related clinical outcomes metrics, in comparison with standard care. A variety of outcomes related to deterioration detection were anticipated, and therefore searches were not limited by outcome. Any outcome related to the detection of deterioration was included as a primary outcome for this review. Speci c outcomes reported in the included studies were: intensive care unit (ICU) transfers; and rapid response team activation or cardiac calls...
A variety of other complications related to clinical deterioration were reported and included in the meta-analysis, from minor (for example fainting, or shortness of breath) to major (such as life-threatening events). A separate analysis was then conducted for the studies separately reporting major complications; the Clavien-Dindo system [17] was applied to postoperative complications in the included studies. This system grades complications from I (deviation from usual recovery not requiring intervention) to V (patient death). To be included in our major complication meta-analysis we included complications de ned by a Clavien-Dindo grade of > II (8,18). Patient death and ICU transfer were not included in this analysis and were assessed separately.
Outcomes reported in less than 3 studies were not included in the meta-analysis and were instead narratively described such as time to antibiotic administration in case of sepsis and number of the National Early Warning Score (NEWS) measurements.

Secondary Outcomes
Secondary outcomes in the meta-analysis included in-hospital mortality and hospital length of stay. Further secondary patient outcomes were reported in the narrative analysis, such as 30-day readmission rates and time to post-operative mobilisation Exploratory Outcomes Exploratory outcomes included the alerting systems used, implementation and iterations in clinical practice. This included type of early warning score, alert thresholds used for each vital sign or overall EWS and other relevant alert parameters/information, where available.
Clinical trial registry searches were also conducted. For included studies that were registered, a comparison was made between the details in registration and report of the study. Registered studies eligible for inclusion but without published results were also narratively reported.

Population and interventions
Complete inclusion and exclusion criteria are available in the published protocol [16]. We included any studies conducted in hospitalised patients, excluding studies conducted in healthy volunteers or non-hospitalised patients.
Studies were eligible for inclusion if they used an ambulatory monitoring system (with or without standard care) in comparison with standard care. An ambulatory monitoring system was de ned as a wearable system monitoring at least one vital sign (heart rate, respiratory rate, temperature, blood pressure or oxygen saturation) sampled continuously at a high rate (e.g. under a minute) or low rate (e.g. every 5 minutes) and where measurements did not require frequent manual input from clinical staff. For comparator we considered any type of standard care for vital sign monitoring, as de ned in the protocol [16]. Study types Studies with the following designs were considered for inclusion: randomised controlled trials (RCTs), cluster RCTs, interventional studies, observational studies (including case-control and before-after studies), and pilot studies. Retrospective studies that complied with the proposed outcomes and eligibility criteria, and unpublished (grey) literature, were also considered. Clinical trials and prospective studies registered up to 10th September 2020 in ClinicalTrials.gov via https://clinicaltrials.gov/ and ISRCTN via https://www.isrctn.com/ were also identi ed. Search details in Appendix 3.

Literature Search and Selection of Studies
Search of unpublished grey literature and pre-print servers was also conducted manually (search details in Appendix 4) and additional studies published in these servers up to 16th December 2020 were identi ed.
Titles and abstracts of all potentially relevant articles were independently reviewed for possible inclusion by two authors (CA, CB). The full text of any citation considered potentially relevant by any reviewer was retrieved. The degree of interrater agreement for study selection was determined by using kappa, with standard de nitions for poor (< 0.20), fair (0.21 to 0.40), moderate (0.41 to 0.60), good (0.61 to 0.80), and very good (0.81 to 1.00) agreement [18,19]. The included abstracts full-texts were assessed for eligibility and disagreements resolved by discussion between the 2 review authors; if no agreement could be reached, a third author was consulted (SV). The full selection process is outlined in the published protocol [16].

Data collection and extraction
Two reviewers (CA and CB) extracted the data independently from the included studies. Disagreements were resolved by discussion between the 2 review authors. When required, this was also discussed with a third author (MS) and a statistician (SG).
The following data were extracted for each study, where available: author list, country, date published, registration number, aim, design, setting and population, recruitment start and end dates, ethical approval and informed consent information, eligibility criteria, intervention description, included devices, period of device wear, vital signs measured by devices, frequency of ambulatory data availability, comparator type, EWS and frequency of manual measurements, sample size, demographics (e.g. age, gender, BMI, etc.), other clinical characteristics (e.g. type of admission, American Society of Anaesthesiologists, ASA, score, etc.), deterioration detection and related clinical outcomes summary data, total/median monitoring hours, alerting system information (e.g. thresholds and alarms description), study limitations, device FDA/CE mark information, funding and con ict of interest information.

Risk of Bias of Individual Studies
Four tools, selected based on study design, were used to assess risk of bias. For randomised controlled trials the Cochrane risk of bias tool (RoB2) was used [20]; for non-randomised studies, the Newcastle Ottawa Scale (NOS) [21] and the "Risk Of Bias In Non-randomised Studies -of Interventions" (ROBINS-I) were applied [22]; and, in addition, the Mixed Methods Appraisal tool (MMAT) [23] was used for all studies. This was a change from the original protocol [16] as the Jadad scale was replaced by the ROBINS-I for assessment of non-randomised studies, as we found it more comparable with the ROB2 tool used for included RCTs. Two reviewers (CA and CB) independently appraised each study and disagreements were solved by discussion until consensus was reached with a third reviewer (SV).

Data Analysis
All outcomes with results from at least three studies were considered for the meta-analysis. Outcomes with data from less than three studies were not included in the meta-analysis but reported in a narrative synthesis. Data preparation and meta-analysis Review Manager 5.4.1 (The Cochrane Collaboration, Oxford, England [24]) was used to calculate pooled risk ratios (RRs) for dichotomous outcomes and pooled weighted mean differences (WMDs) for continuous outcomes, and respective 95% con dence intervals (CIs). Continuous variables are expressed as mean (SD). Due to differences in design between included studies, we used random-effects meta-analysis and the TAU2 statistic, and respective signi cance level was calculated [25]. We assessed heterogeneity among trials by using I 2 (the percentage of total variability across studies attributable to heterogeneity rather than to chance) and used published guidelines for interpretation [24].
One before-and-after study compared the AMS group with a before period in the same unit and a different unit (both before and during) [26]. For the metaanalysis, we limited data to that reported from the same unit to minimising confounding. Outcomes for this study were also presented per 1000 discharges. As the authors provided the total number of discharges, the actual event numbers were calculated for inclusion in our analysis [26]. Similarly, another included study presented the hospital length of stay (LoS) in hours [27], this was converted to days for the analysis. In a further study, the authors presented LoS in median (Interquartile range, IQR), which was converted to mean (standard deviation, SD). A normal distribution of the values was assumed to make this conversion, as per Cochrane guidance [28].
Finally, in one study complication data was presented as the number of events rather than the number of patients suffering a complication [29,30]. A formal data request to the principal investigator was made to acquire the data in the correct format, and this was used in the meta-analysis. Narrative analysis Alerting thresholds, methods and other alarm information was extracted from the included studies, where available, and narratively reported. For analysis of study registration, the proportion of registered studies that were published, and both the dates of trial registration and publication of results were reported. We also explored registered versus published primary and secondary outcomes. Principal Investigators for the included and registered studies were contacted for further information, as required. Body of evidence summary A body of evidence summary is provided in Appendix 5, using the GRADEpro software [31].

Study selection
After removal of duplicates, 8706 studies were identi ed. Following title and abstract review 51 full texts remained, of which 10 met the inclusion criteria ( Fig.   1). Four studies appeared to meet the inclusion criteria but were excluded at full-text review: two studies were excluded for not reporting a subset of their analysis for the patients using the AMS [32,33]; one was excluded after con rming with the author that the device was not ambulatory at the time of the study [34], and another did not have a comparator group [35]. A total of 4433 patients were included in these studies.

Study Characteristics
Devices A variety of devices were used in the studies included in this review: four studies used the VisiMobile (Sotera Visi Mobile, San Diego, California) [26, 36-38], with two studies also adding the HealthPatch (Vital Connect, Campbell, CA, USA) [37,38]; two studies used Sensium Vitals (Sensium, Abingdon, United Kingdom) [29,30]. The remaining four studies used different devices, including the Patient Status Engine (Isansys Lifecare Ltd.) [27], the Auricall monitoring system [39], the Avant-4100 (Nonin) [40] and the Monica Novii wireless patch system (General Electric Company, Milwaukee, WI) [41]. Included studies and outcomes Table 1 Characteristics of the included studies. AMS: ambulatory monitoring system, BP: blood pressure, CEWS: centile-based early warning score, ESS: e cacy saf score, EWS: early warning score, HR: heart rate, NEWS: national early warning score, PSE: Patient Status Engine, RR: respiratory rate, SpO2: peripheral oxyg saturation, T: temperature. *Same patients. Weenk  Of the ten studies identi ed, seven were included in the meta-analysis with a total of 4127 patients. These included two RCTs [27,30], one cluster RCT [29], and four before-and-after observational studies [26,36,39,40]. Three further studies (RCTs) were included in the narrative synthesis [37,38,41], including a total of 306 further patients. Details of the included studies are presented in Tables 1 and 2.
The majority of the included studies implemented the AMS in post-surgical patients. Four studies also reported the patient American Society of Anaesthesiologists (ASA) score for preoperative functional status [42], with a median ASA score of 2 ("Patient has mild systemic disease") in three studies [29,30,43] and 3 ("Patient has severe systemic disease that is not incapacitating.") in one [27].
Reviewers achieved a fair level of agreement (kappa: 0.348; 95% CI 0.285 to 0.482) for study inclusion. There were no major disagreements between reviewers regarding data extraction, study quality or bias assessments. Studies not included in the meta-analysis were narratively explored (Tables 1 and 2). Two papers reported results from one RCT, comparing two devices (HealthPatch and VisiMobile) with nurse measurements [37,38]. However, they have not included the third group (control) in the analysis and not assessed any clinical outcomes, mostly exploring factors related to deterioration detection, failing to provide su cient data to include in the meta-analysis. In the rst paper from this RCT, the authors report that both HealthPatch and VisiMobile modi ed early warning scores (MEWS) were higher than the nurse measured MEWS, mostly due to RR measurements differences [38]. In the second paper (the full RCT) the authors identi ed positive and negative effects as well as barriers and facilitators for the use of these devices, such as the impact of AMS in a shorter length of stay and prevention of ICU admissions, additionally, a total of 17 patients, 2 relatives and 17 healthcare professionals reported to be expecting earlier deterioration detection using these wearables [37].
Another RCT evaluated wireless external fetal electrocardiography versus standard external monitoring [41]. We were unable to include this study in the metaanalysis as (1) the primary outcome of the study was the percentage of interpretable fetal HR data, (2) the population of interest is very different from the remaining studies and (3) the clinical outcomes analysed also differed (eg. length of labour, fetal Apgar score, etc.). Considering this, their results demonstrated no differences in maternal or neonatal clinical outcomes between groups. However, results did suggest an increased acceptance by patients and staff, with satisfaction scores signi cantly higher when compared to the standard monitor [41]. Included studies registration Details of the clinical trials search are shown in Appendix 6. Of the ten included studies in this review, only seven were registered (most retrospectively, as per Appendix 7). Within these, all primary outcomes stated in the registration were reported in the main paper, as well as most of the secondary outcomes.

Study quality and risk of bias
The overall quality of included studies was moderate with some bias to take into account, as per Figs. 2 and 3. For the included RCTs, using the ROB2, two were identi ed as at "low risk" of bias [30,41] and a further three were assessed as raising "some concerns" [27,37], including the cluster RCT [29]. The risk of bias, assessed by the ROBINS-I was "moderate" for all before-and-after studies [26,36,39,40]. See Appendix 8 for further details. The results of the bias assessment did not in uence inclusion in the meta-analyses.

Primary outcomes
In total, data from seven studies were included in the meta-analysis of primary outcomes related to deterioration detection, analysed separately according to the three reported deterioration outcomes -ICU transfers, rapid response or cardiac arrest activation, and complications.

ICU transfers
A total of ve studies reported ICU transfers and were included in this meta-analysis (data from 3565 patients, 1898 in the AMS group) [26, 29,30,36,40].
Pooling of data indicated that use of AMS did reduce ICU transfer (RR, 0.87; 95% CI 0.66 to 1.15), but not statistically signi cantly (p = 0.32) (Fig. 4). Rapid response or cardiac arrest activation For this outcome, two before-and-after studies reporting rapid response team activation and another study reporting cardiac arrest calls were included (with data from 3214 patients, 1698 in the AMS group, Fig. 5 A total of ve studies reported data on complication outcomes classed by the Clavien-Dindo system as grade I or II(with data from 1752 patients, 837 in the AMS group, Fig. 6). indicating the AMS group had a reduced risk of complications (RR, 0.77; 95% CI 0.44 to 1.32) however without statistical signi cance (p = 0.34) and with high heterogeneity between studies (I 2 = 93%).
For the major complications (Fig. 7), we included 3 studies (with data from 546 patients, 296 in the AMS group) indicating the AMS group had reduced risk of major complications (RR, 0.55; 95% CI 0.24 to 1.30) however, with no statistical signi cance (p = 0.17).
Other deterioration detection outcomes not included in the meta-analysis A few of the included studies also explored other deterioration detection outcomes, but in insu cient numbers to allow a meta-analysis. One cluster RCT [29] and one RCT [30] from the same research group compared the time to antibiotic administration in case of sepsis in the AMS group against the control group, nding this statistically insigni cant in both studies (656.0 (95% CI 431.7-820.3) vs 1012.8 (95% CI 425.0-1600.6) minutes [29] and 551 (95% CI 296-805) vs 527 (95% CI 199-856)) [30].

Secondary outcomes
The two secondary outcomes of in-hospital mortality and hospital length of stay were also meta-analysed.

In-hospital mortality
For the outcome of in-hospital mortality, we included six studies (with data from 3760 patients, 1994 in the AMS group, Fig. 8) [27], that the authors defend being a result of the increased monitoring in the AMS group, facilitating pain and oxygen management of those patients, and promoting earlier mobilisation [27]. This study also  (18,48). Despite this, two patients withdrew from the study due to "too many false alerts" (18). In two other studies, the authors just discussed the intention was to improve the rate of true positives and reduce false negatives/false alarms [37,38].
Most alerting thresholds were pre-set and individualised as required, alerted through the central station and/or nurse mobile/pager/PDA, with the majority using audio alerts (Table 4). In one study the authors used a single risk score calculated from all vital signs (VSI) and based on modelling from a previous patient dataset, creating an alert when the VSI score was above the threshold for more than 4 out of 5 minutes [40]. Alerting parameters from the included studies are explored in Table 4.
In one study's nal version of the alerting system, hypotension, bradycardia and hypoxemia were tolerated for shorter periods than tachycardia or hypertension, unless the tachycardia resulted in hypotension. Additionally, the majority of the alerts in this nal iteration were due to SpO2 (0.97 APDs) [26].
Another study found that the most accurate vital sign parameter was systolic blood pressure, which had a positive predictive value (PPV) of 97%, followed by high respiratory rate (PPV of 85%) and low SpO 2 (PPV of 76%), indicating high sensitivity and reliability and a low false alarm rate.
[36] A total of sixteen registrations were identi ed in the search and screened for eligibility. Six were excluded and six registrations refer to the included studies. A further four registrations were deemed potentially eligible to be included in our review and meta-analysis (Table 3). A registered cluster RCT [45] aimed to develop a two-tiered monitoring system to improve the care of patients at risk of clinical deterioration in general hospital wards. This registered study also included a subset of patients using wireless devices [45]. However, although the main results are published [32], no data were reported on the impact of wireless devices on the subset population and we were unable to make contact with the Principal Investigator to clarify publication status and request this subset of the AMS group data. A further registration [46] before-and-after study was potentially eligible and although the main results are published [33] a subset of patients (278) used at least one cableless sensor, but the author con rmed there were no data available on outcomes for this sub-set of wirelessly monitored patients [33]. The other two registrations did not publish their results at the time of our systematic literature search. The Principal Investigator of one prospective, observational cohort study con rmed the study results have been submitted and is under peer review [47]. We were unable to contact the Principal Investigator to clarify the status of the other study [48]. . This review also provides some limited indication that AMS may reduce in-hospital mortality and length of stay, but again without strong statistical signi cance supporting these ndings.
Although our review focused speci cally on ambulatory monitoring devices, our results are in accordance with a previous systematic review with a focus on the clinical impact of a broader range of multi-parameter continuous non-invasive monitoring of vital signs in non-intensive care unit patients. In this previous review, the authors included all non-invasive devices (including wired and static bed monitors such as EarlySense) and also found a trend towards reduced ICU transfers, RRT activations and hospital length of stay. It also suggested reduced hospital mortality in patients using these devices [11]. Our study also updates their review with the inclusion of an additional four more recent studies.
All the studies included in this review were conducted in a non-ICU environment (mostly in surgical patients) and most comparators were standard intermittent vital sign monitoring with the use of local EWS. Previous evidence suggests that focus on non-critical care settings is due to AMS being unable to replace the continuous monitoring commonly used for high dependency patients. Instead, AMS offers an intermediate level of monitoring between continuous high dependency monitoring and intermittent manual measurements, with the potential to facilitate early deterioration detection in high-risk patients (e.g. post-ICU) [2]. In addition, a recent study in the paediatric population concluded that wireless monitoring is feasible and can identify more deteriorations. The authors suggest that by using this in combination with a paediatrics early warning (PEW), some life-threatening events may be prevented [35].
One study included in this review reported patients', relatives' and healthcare professionals' perceptions of the use of AMS in the general ward, and found agreement between all interviewed groups that AMS could facilitate earlier deterioration detection and improve patient safety without posing a barrier to mobility, as well as reduce staff workload and hospital costs [37], agreeing with previous evidence [49][50][51] and reinforcing the direction of our ndings.
To better understand the used AMS, we have aggregated available information, of the alerting methods and thresholds of the included studies. As most studies used audio alerts, system iterations seemed to focus on reducing the rate of false alerts by adjusting/individualising each vital sign and/or overall score to avoid alarm fatigue from clinical staff. This has been previously discussed as an important factor for the successful deployment of monitoring technology [52][53][54][55]. The exception was one study that used visual warnings alone, resulting in increased NEWS measurements in patients using AMS without clinical staff being aware of the potential deterioration. They identi ed three patients for whom AMS alarmed between two intermittent measurements and who were later diagnosed with a pneumonia, atrial brillation and an anastomotic leakage [38]; in the same study, the authors also explored the delay between high MEWS measured by a device and next regular MEWS measurement by a nurse, ranging from 0 up to 10 hours, and varying between day and night [38].
Finally, a review of known registries was conducted to address the issue of publication bias, nding that the majority of the included studies were registered either in ClinicalTrials.gov or ISRCTN databases and all registered primary outcomes were reported as well as most of the secondary (Appendix 7). However, most studies were registered retrospectively, rather than prospectively. We did, however, identify four studies that might have contributed to this systematic review by either performing a subgroup analysis for patients using ambulatory devices [33,45] and by publishing their results as registered in another two [47,48]. Our registry search allowed to highlight this under-reported and non-published evidence that could have potentially impacted our results and contributed to the meta-analysis and overall body of evidence in this eld.

Study limitations
There were some limitations to this review. The number of studies included was limited and used a variety of designs, populations, outcomes, medical devices/systems, EWS and alerting thresholds. Our meta-analysis only included two RCTs (one being a pilot) and one pilot cluster RCT, all of which had small sample sizes which reduced the probability of signi cant difference in outcomes, despite large effect sizes. In contrast, the included before-and-after studies had larger sample sizes but increased bias and quality limitations, again posing a barrier to any signi cant conclusion. This re ects the emerging nature of this area of research and highlights the need for a large, multicentre RCT that tests whether AMS may be bene cial for early deterioration detection and related clinical outcomes in hospitalised patients.
One of the studies [29] had an extra non-randomised bay in their exploratory analysis. We did not include these data in our meta-analysis in accordance with Cochrane guidelines, so only the randomised groups were considered [56]. Similarly, another study compared the AMS group with a period before in the same unit and a period during and before in an alternative unit. In this case, we only used the data comparing the before period in the same unit, to minimise selection bias [56]. For RRT and cardiac team call analysis, it is important to note that Weller and colleagues study accounted for 96.7% of the weight on these results [26]. In addition, studies exploring RRT and cardiac team calls were combined in the meta-analysis, and therefore results might differ if analysed separately.
As previously reported, bias may be present as clinical staff is aware of the AMS use in their patients. However, the practicalities of blinding in AMS studies may not be feasible and may potentially have a counterproductive impact on clinical outcomes [11].

Future research
The results of this systematic review suggest the use of AMS might promote early deterioration detection and improve clinical outcomes. However, these ndings are limited by the reduced sample size of the included RCTs and reduced methodological quality of the before-and-after studies. One of the included studies [30] conducted a subgroup analysis exploring clinical outcomes in high-risk participants. This further highlighted the impact of the AMS use in patients at higher risk of deterioration and further RCTs should consider including a similar analysis of their results.
Remote monitoring systems are only bene cial to the clinical staff if they are easy to use and clinicians understand the potential bene t for patient clinical outcomes [57]. This review is part of a wider phased project, the virtual High Dependency Unit (vHDU) study. So far, we have selected [58] and tested AMDs [59] to be integrated in a nal AMS, that will be tested in a pilot and then a full multicentre RCT in the near future.

Conclusions
Our systematic review suggests that AMS may have a positive impact on early deterioration detection and associated clinical outcomes. Although no results were signi cant, we did nd a trend towards fewer ICU admissions, RRT/cardiac arrest team calls, major and overall complications, in-hospital mortality and hospital length of stay in patients using the AMS against standard care. The included studies were limited by a reduced number of RCTs with a small sample size and before-and-after studies with moderate bias. This review highlights the need for bigger and more rigorous RCTs to support our meta-analysis ndings.
Additionally, our narrative ndings suggest that alarms should be adjusted to minimise false alerts and thus support prompt clinical action to deterioration events. CA designed the study. NT conducted databases search. PW and LT made nal approval of the study protocol. CA and CB selected studies for inclusion, SV resolved any disagreements. CA and CB extracted the data from studies, SG and MS reviewed data extraction. CA and CB conducted quality assessment, SV resolved disagreements. CA conducted statistical analysis with SG supervision. CA wrote the rst draft of the manuscript with SV support. All authors contributed and approved the nal manuscript.

Figure 2
Risk of bias summary: review authors' judgements about each risk of bias item for each included study.  Meta-analysis forest plots comparing AMS and standard care ICU transfer risk ratio. AMS: ambulatory monitoring systems, CI: con dence intervals, M-H: Mantel-Haenszel, RCT: randomised controlled trial.  Meta-analysis forest plots comparing AMS and standard care in-hospital mortality risk ratio. AMS: ambulatory monitoring systems, CI: con dence intervals, M-H: Mantel-Haenszel, RCT: randomised controlled trial.

Figure 9
Meta-analysis forest plots comparing AMS and standard care hospital length of stay mean difference. AMS: ambulatory monitoring systems, CI: con dence intervals, IV: inverse variance, RCT: randomised controlled trial Supplementary Files