We first identified adult patients (aged ≥18 years) admitted to a tertiary care center (Atrium Health Wake Forest Baptist Medical Center main campus) over 2 years (between January 1, 2018, and December 31, 2019) under a primary medical service.

Primary medical services included hospitalist, general internal medicine, and family medicine. Provider staffing varied by service but included physician attendings, advanced practice providers, and medical residents. Primary medical services represented the typical breadth of patients admitted to hospital medicine services. Admission diagnoses were not included in propensity matching. Patients having any surgery during admission were excluded from analysis. Race and ethnicity were self-described by patients on admission then retrospectively collected with other demographics from the electronic health record (EHR).

Regarding human subject research ethics review, our study was a secondary analysis and did not actively involve nor enroll human subjects. Our study and a priori defined statistical analysis plan were approved by the Wake Forest University Health Sciences Institutional Review Board (IRB; approval number IRB00083520). Informed consent from human subjects was not obtained for this secondary analysis. Primary consent for data collection was allowed by the IRB without additional consent. Privacy and confidentiality protection were maintained using de-identified data for analysis and using only data needed for the analysis. Data were stored in a secure location per IRB requirements.

Patients admitted to the medical wards either received (1) VS monitoring with a continuous monitoring device, ViSi Mobile, (Sotera Wireless Inc), or (2) standard intermittent VS monitoring (every 4-8 hours). Briefly, the ViSi Mobile system is a portable wrist-mounted system that received clearance from the Food and Drug Administration in 2012. These ViSi devices were purchased and implemented by the Wake Forest Baptist Hospital system in 2015. Today, ViSi monitors are on several different medical and surgical units; however, these devices are not available on all units. If units do not have ViSi monitors or patients decline to wear the monitor, then intermittent monitoring is used.

Hospital units with ViSi devices are staffed by the same number of personnel as those not doing this monitoring. Nurse-to-patient ratios on our floors range from 1:3 to 1:6. For this analysis, patients were not randomized into continuous or intermittent VS monitoring. The type of monitoring was dependent upon the floor a patient was admitted to, which was dependent on bed availability and not patient clinical status. Based on clinical experience, there would not have been a significant crossover between groups; that is, once a patient arrived on the floor and was allocated to either continuous or intermittent VS monitoring, they would then most likely continue that strategy until discharge—unless, for example, a patient was transferred to a different unit and received a different type of monitoring strategy there.

The ViSi device records oxygen saturation (SpO₂), HR, respiratory rate (RR), continuous noninvasive BP, and either 3-lead or 5-lead electrocardiogram. Continuous BP measurements are estimated from pulse arrival time, specifically the time that elapses between the R wave being detected and the arrival of the resulting pulse at the SpO₂ finger sensor. According to the manufacturer, the estimated maximal mean error for BP measurements is 5 mm Hg with a standard deviation ≤8 mm Hg [23].  

ViSi monitors were calibrated daily with the oscillometric brachial cuff method and connected to the hospital’s wireless network, and measurements were distributed to the central nurse’s station for that unit and continuously displayed. The ViSi monitoring system measures VS nearly continuously, and these data are transmitted to the central nurse’s station for each respective unit; however, these data are not transmitted nor recorded directly in the EHR. That is, nurses or nursing assistants must still see the patient and collect VS from the device or the monitoring station remotely and then manually document (chart) these vitals in the EHR. No matter what, the VS assessment includes direct patient visualization to measure and record patient alertness. The VS data not recorded in the EHR are temporarily stored on the ViSi manufacturer’s cloud for a set period of time and are then destroyed. Unfortunately, we were only able to acquire and evaluate the EHR data.

VS abnormalities exceeding established thresholds for the local hospital system generated alerts at the floor’s central nursing station and to the nurses’ hospital-supplied phones (ASCOM). Alarms not addressed by the primary nurses were escalated to other floor nurses, then to the unit manager. Time to escalation depended on the VS and the relative degree of abnormality. For example, SpO₂ abnormality alerts, if not acknowledged, were quickly escalated over minutes.

EWS at our institution are based on the patient’s weighted VS of HR, RR, SpO₂, temperature, BP, rate and percentage of inspired oxygen, and level of awareness determined by the patient’s nurse [24]. EWS are calculated by the EHR every 4 hours based on data input from ViSi monitors, which can be collected from the central display or at the bedside, or those collected by intermittent monitoring. Patients, regardless of monitoring strategy, must be observed at bedside to determine level of awareness for EWS calculation. Total scores range from 0 to 20. Ad hoc EWS may be generated by the clinical team in response to changes in a patient’s clinical status through either monitoring strategy. For example, if abnormal VS are detected by ViSi monitoring and an alarm is generated at the central nurse’s station, the clinical team may assess the patient and generate an updated EWS. An RRT activation is triggered if the EWS is ≥8.

Our primary outcome was the risk of a composite of in-hospital mortality or ICU admission during the hospital encounter. Secondary outcomes were the individual components of the primary outcome (in-hospital mortality and ICU admission), as well as HF, MI, AKI, and RRT activations throughout the hospitalization. Exploratory outcomes included hospital length of stay (LOS), ICU LOS, and median EWS values for the entire LOS. Diagnoses of HF, MI, and AKI were identified using International Classification of Diseases, Tenth Revision billing codes available in the EHR. 

Statistical analysis was performed in R v3.6.1 (R Foundation for Statistical Computing) using RStudio environment v1.1.456. The participants were first stratified into continuous VS monitoring and contemporaneous controls as intermittent VS monitoring groups. Sequential random nearest neighbor matching with a propensity score ratio of 1:3 and with a caliper of 0.25 [25,26] was used to balance the potential factors affecting the outcome including selection bias. To create propensity score–matched pairs, we performed random 1-to-3 matching using the MatchIt package in R, in which the control variables in our study were race, gender, ethnicity, age, quarter-year of admission, BMI, Charlson Comorbidity Index (CCI), patient’s primary insurance type, and source of admission [27]. The variable selection for propensity score matching was based on feedback from clinical expertise. To adjust for the effect of time and variability of the continuous wireless monitoring device use, the quarter-year of admission was used as a covariate in the model. Standardized differences were used as indicators of intergroup balance. If the standardized differences were <10%, the covariates between the 2 groups were considered well-balanced.

For summary statistics, categorical variables were described as numbers and percentages, and non-normally distributed continuous variables were described as median and IQRs.

Primary, secondary, and exploratory outcomes were compared between groups and subgroups. Multivariable logistic regression analysis was performed to detect differences in outcomes between groups. Initially, restricted cubic splines were included with age and CCI to relax the linearity assumptions of the multivariable logistic regression model. As the nonlinear components of age and CCI were not statistically significant, the splines were removed from the variables. The multivariable logistic regression adjusted model was controlled for race, gender, ethnicity, age, BMI, patient’s primary insurance type, source of admission, and CCI. The P value for significance level within the primary and secondary outcome analyses was set at .004 for each group of analyses after Bonferroni correction for multiple comparisons.  

To test the influence of unobserved confounding, we used an E-value analysis to quantify the potential implications of unmeasured confounders. E-value is defined as the minimum strength of an association on the risk ratio scale that an unmeasured confounder would need to have with both the exposure and the outcome, conditional on the measured covariates, to fully explain away a specific exposure-outcome association. 

We further performed multivariate logistic regression analysis for the composite outcome on the whole study cohort (n=7971) to assess the effect of continuous monitoring sensitivity after controlling for race, gender, ethnicity, age, BMI, and CCI as a sensitivity analysis.

In this retrospective propensity-matched analysis, the standard of care intermittent VS monitoring of adult medical ward patients was associated with nearly 3-fold greater odds of ICU admission or in-hospital mortality as compared with continuous VS monitoring. Our work is novel in that this is, to the best of our knowledge, the first propensity-matched study of continuous VS monitoring in the medical ward. Primary medical admissions are inherently different from surgical ones. Medical patients often have multiple acute problems related to single or multiple intersected disease processes.

Our findings corroborate several before-and-after studies mainly conducted with surgical cohorts and are additionally supported by a recently published propensity-matched study by our group in a primary surgical population [28-30]. We hypothesized that, through earlier detection with continuous VS monitoring, patients were able to receive clinical attention sooner, thereby precluding severe clinical deterioration requiring ICU admission. This hypothesis is supported by similar RRT activation rates but greater odds of ICU admission of patients receiving intermittent VS monitoring.

We report a significant difference in favor of continuous monitoring for the primary composite outcome; however, this composite is only from ICU admissions. When compared with a recently published national database study of unplanned ICU admissions of patients who received standard of care intermittent monitoring, we had a greater percentage of patients in our intermittent monitoring arm transfer to the ICU (2.5% vs 5.8%) [31]. Baseline characteristics, including median CCI, were similar; however, the national study included patients with COVID-19 (2021), while our study examined a prepandemic cohort (2018-2019). Moreover, hospitals in the national database varied significantly from ours in terms of size, teaching status, and regional location, which could have influenced outcomes.

Our study was limited primarily by its retrospective approach. First, although propensity matching allowed us to match patients on baseline covariates, we were unable to match based on admission diagnosis or ongoing changes in a patient’s hospital course. We were unable to match on admission diagnosis due to limitations in our EHR data (eg, admission diagnosis missing, inaccurate, or incorrect). Changes over the hospital course occurred after receiving the intervention (continuous versus intermittent monitoring); therefore, the changes were not used to match patients. The inability to match admission diagnosis is a major limitation, but we feel this is offset by our hospital’s triage system by bed availability, which is agnostic to monitoring strategy. The use of continuous wireless VS monitoring or intermittent VS monitoring was not based on patient condition, but it is possible that, over time, patients admitted to continuous VS monitoring units may have represented a subset of clinically different patients from their intermittent VS monitoring counterparts. That is, patients receiving intermittent VS monitoring may have initially been sicker, or more continuous wireless VS monitoring use may have been deployed over a time course to a healthier group. To account for these differences, cohorts were matched on a combination of baseline covariates present on admission along with every quarter-year of admission to adjust for time variance of continuous wireless VS monitoring use throughout our large hospital system. The difference between the number of patients receiving continuous VS monitoring in 2018 compared with 2019 is related to an issue that occurred with the continuous VS monitoring server. Our approach with propensity matching was set at a caliper of 0.25, and the drop-off for continuously monitored patients does demonstrate that patient groups may be different at baseline within the defined caliper. Most of the imbalance was due to the time of year of admission and CCI, which was matched adequately. An increase in the caliper from 0.25 to 0.35 did not show any significant change in the matched cohort number. Second, we relied on reported rates of conditions from discharge summaries and administrative codes, which have their inherent flaws.

Third, our large administrative data set does not have granular information that would allow a deeper understanding of the mechanisms of favorable outcomes with continuous VS monitoring. This includes the number of and reasons for alarms and false alarms as well as the VS captured by the continuous devices. Similarly, we had limited information on EWS values at the time of RRT activation but did include median, minimum, and maximum EWS and absolute numbers of RRT activations for the entire LOS for both groups. A smaller pilot from the same institution and devices showed a rate of 2.3 alarms per patient per day and a significant decrease in RRT calls, from 189 to 158 per 1000 discharges after implementation using a before-and-after study design [21].

Interesting future outcome research could include whether continuous VS monitoring was associated with changes in medical management, including the timing of antibiotics, discontinuation of supplemental oxygen, and antihypertensive prescriptions. Future work should consider the limitations of this study in order to strengthen the approach, including admission diagnosis matching, granular VS capture, and comparisons between monitoring strategies.

The ability to continuously monitor patient VS outside of the critical care unit or emergency department using a wearable and wireless device represents a significant technological advancement. However, the quality of evidence for the benefit of continuous VS monitoring is low to moderate, and most studies have been performed in the postoperative setting with before-and-after comparisons. This is the first propensity-matched study using a large data set with contemporaneous controls and a focus on patients with primary medical problems. The inherent differences between a primary surgical population, where nearly all continuous postprocedure monitoring research has been conducted, and a primary medical population is paramount to the significance of this work.

In conclusion, patients receiving intermittent VS monitoring had nearly a factor-of-3 increase in the odds of ICU transfer and in-hospital mortality compared with those receiving continuous VS monitoring. Prospective interventional studies are necessary to confirm the impact of continuous VS monitoring on morbidity and mortality.