1 Background
In 2004 the Institute for Healthcare Improvement (IHI) launched the “100,000 Lives Campaign” to significantly reduce morbidity and mortality in the US healthcare system. Six steps were identified, one of which was to deploy rapid response teams at the first sign of patient decline. In 2006, with the initial goal exceeded, IHI introduced a “5 Million Lives Campaign” again recommending rapid response teams as a core component recognizing that ‘failure to rescue’ events are a major cause of mortality in American hospitals.
An international consensus conference suggested that to reduce failure to rescue events a rapid response system (RRS) is necessary. A RRS is composed of four components: an afferent limb (detection and response triggers), an “efferent limb” (technical and human resources brought to the bedside), a quality limb, and an administrative limb [
1]. Studies investigating in-hospital mortality have shown deterioration of vital signs 6–12 h prior to a serious clinical event [
2‐
4]. Yet, despite rising inpatient acuity levels, the standard for routine physiological assessment outside the ICU is once every 4–8 h. The conference identified a core set of parameters which should be continually monitored for the goal of early detection of physiological instability: heart rate (HR), blood pressure, respiratory rate, temperature, pulse oximetry, and level of consciousness [
1].
One of the concerns about implementing multi-parameter continual vital sign monitoring on a general medical or medical-surgical unit is alarm fatigue. The high rate of alarms, and their sometimes limited clinical relevance has been well documented in the operating room [
5] and ICU [
6,
7] and the concern about alarm fatigue is real [
7]. In general medical units, the nurse to patient ratio is lower than in the OR and ICU, so continual monitoring in this environment has the potential for being a significant burden on nurses.
Several methods have been evaluated to reduce the number of non-actionable alarms during continual monitoring of patients. One successful approach is to combine alarm thresholds with annunciation delays (a delay between when an alarm threshold has been crossed and when the alert is sounded or displayed) [
8‐
12]. In most cases alarm thresholds are set based upon knowledge of and experience with the vital sign measures. Recently, work has been done to develop evidence-based methods for determining alarm limits. Burgess et al. [
11] established a database of HR and respiration rate (RR) measurements from 317 patients (18,737 h) in a general care unit, with no adverse events. Modeling was done with this database to predict the alarm rate for different alarm limits, with the goal of reducing the number of false positive alarms. Welch [
12] describes a method for reducing SpO
2 alarms based on creating a large database of measurements. The database is used to predict the alarm rate based upon combinations of alarm thresholds and annunciation delays. The current study extends Welch’s methodology [
12] to also include HR, RR, systolic, diastolic and mean blood pressure (SBP, DBP, MAP), by evaluating a method for optimizing alarm rates for continual multi-parameter monitoring in general care units.
2 Methods
2.1 Vital sign data collection
Vital sign data were collected throughout the day and night with the ViSi Mobile System, an on-body, multi-parameter monitoring platform capable of continual measurement and display of core vital signs including ECG, RR, HR, continual NIBP (cNIBP), pulse oximetry (SpO
2), pulse rate, and skin temperature (Sotera Wireless Inc., San Diego, CA, USA) [
13]. Data capture included waveforms (500 samples/s) and numeric data displayed as a 3 s moving average.
The ViSi Mobile System transmitted patient data via the existing wireless network in the hospital. De-identified patient data were uploaded through a secure link to a private cloud at the end of each patient’s monitoring session. The numeric values in the cloud-hosted database were then used to simulate the total alarm burden associated with simultaneous monitoring of all key vital signs.
2.2 Alarm rate simulation and independent hospital evaluation
A range of potential alarm thresholds were chosen by the authors and used to model the impact of annunciation delays on the resulting total alarm rate. Ranges for the alarm thresholds used in the simulations were chosen based upon the distribution of vital sign values in the database (Supplemental Data). Adverse events in general care units are rare, so in a large population it can be assumed that alarm thresholds should be set near the tail ends of each distribution. Tables were constructed for each vital sign showing the projected number of alarms for each combination of threshold and annunciation delay. Specific alarm thresholds and annunciation delays were selected for each vital sign and the total alarm rates (number of alarms/patient/day) were calculated. A 10 day evaluation was conducted at an independent hospital, not included in the cloud-hosted database, over the period January 20, 2015–January 30, 2015. Continual SpO2, HR, RR, cNIBP data were collected for these patients and alarm rates were calculated using the same thresholds and annunciation delays as above. The alarm rate calculated for the independent hospital was compared with the simulation done with the data in the cloud-hosted database.
2.3 Database comparisons
Since alarm threshold values were determined from the population histograms of data in the cloud-hosted database, we assessed general applicability of the method by comparing these distributions to data from previously published databases. One comparison was made with the 1.15 million individual vital sign determinations from 27,722 patients reported by Bleyer et al. [
14], collected intermittently for patients on all non-intensive care unit and all non-intermediate care unit floors at a single institution. The data to re-plot the histograms for the comparison were obtained from the online supplement [
14]. A second comparison was made to data from Tarassenko et al. [
15]; 64,622 h of vital sign measurements from 863 patients. These data were collected from continual five-parameter monitoring in med-surg patients at one hospital in the US and one hospital in the UK. To re-plot the histograms the data were read digitally from Figure 1 in the Tarassenko paper [
15] using WebPlotDigitizer Version 3.4 (A. Rohatgi 2014,
http://arohatgi.info/WebPlotDigitizer). For comparison, all histograms were normalized so the total area under the curve was 1.
4 Discussion
This study demonstrated the use of a large database for simulating the effect of alarm thresholds and annunciation delays on the total alarm burden we anticipate might be experienced by nurses doing multi-parameter patient monitoring in a general care unit. The cloud-hosted database was used to perform numerous simulations over a large range of alarm thresholds and annunciation delays. When the total alarm burden was estimated using specific threshold and annunciation delays for each vital sign the same total alarm rate was found for both the large database and the independent test hospital, indicating that the cloud-hosted database is a good representation of the patient population in a general care unit.
Low alarm rates can be achieved by selecting thresholds that represent the lowest and highest 0.5–1.0 % of each vital sign distribution (Table
4) and choosing annunciation delays that suppress alerts that resolve themselves quickly (Table
2; Fig.
1). We did not assess if the suppressed alarms were indeed actionable clinical events, but including an annunciation delay is known to eliminate transient and motion artifacts which can be a major source of false alarms [
10].
It has been estimated that there are between 100 and 200 alarms/patient/day in the ICU [
6,
7]. Patients in general care units are presumed to be more stable, and therefore would be expected to have a lower alarm rate. Taenzer et al. [
10] report four alarms/patient/day using a SpO
2-only system. Our results, using a different pulse oximeter demonstrated similar results (3.7 alarms/patient/day for SpO
2). In our study we demonstrate that it is feasible, though not necessarily sufficient, to use population data to assess the impact of alarm thresholds and annunciation delays for all vital signs, giving clinicians control over the number of alerts generated by monitoring systems.
Figure
2 and Table
4 demonstrate that the distribution of vital signs for the general care unit is independent of the monitoring equipment used to make the measurement. Comparison of individual patient values to the distribution may provide guidance to the nurse and aid in their decision making process. Values significantly outside the expected distribution could indicate patient deterioration, equipment malfunction or a patient with unusual physiology. Further patient assessment might lead to the appropriate action: therapy, equipment adjustments (replace dislodged nasal cannula, for example) or individualized alarm settings, respectively for the causes listed above.
4.1 Limitations
This study demonstrated a data-driven methodology for managing the number of alarms when multi-parameter monitoring is done in general care units. While the methodology is generalizable, the results in Fig.
2 and Table
4 and the supplement are specific to general care units and would not be applicable for continual monitoring in the ICU or the OR. The specific values for alarm thresholds and annunciation delays have been determined for continual vital sign monitoring and may not be appropriate for spot check monitoring for general care units, where intermittent monitoring may require higher sensitivity.
The cloud-hosted database represents ten hospitals with data collection over a 9 month period. cNIBP data became more prevalent during the later periods of data collection, hence the percentage of patients with cNIBP measurements in the independent hospital was higher than in the overall database. It is likely that this observation explains the higher rate of cNIBP alarms in the test hospital data. The database continues to grow and is expected in the future to be more representative in predicting the number of cNIBP alarms. Similarly, during the course of the study we made algorithm improvements to the pulse oximeter sensor which lowered the alarm rate for SpO2 and HR for the independent hospital.
This study described one way data can be used to manage alarms and alerts during continual multi-parameter monitoring in general care units. This is in contrast to the ICU setting where there are a large number of patient adverse events and alarm thresholds are optimized for enhanced sensitivity to these events. The work described here was not designed to demonstrate sensitivity of the alarm settings to actionable clinical events, nor was it designed to compare this methodology to other methods of using vital sign data to alert nurses to patient deterioration.
5 Conclusions
This study produced a large cloud-hosted database documenting the distribution of all standard vital sign measurements for 94,575 h of continual patient monitoring in general care units. The cloud-hosted database was used to demonstrate a methodology for selecting alarm thresholds and annunciation delays based on simulations of the total alarm burden that might be expected. The approach was validated with an independent dataset, demonstrating applicability to a new hospital. Similarities in the distribution of vital sign data in the cloud-hosted database with vital sign distributions previously published, suggests that this methodology could also be applied to data collected using other multi-parameter vital sign monitors. This effort is consistent with and directly supportive of the US Joint Commission National Patient Safety Goal 06.01.01.
Acknowledgments
The authors are grateful for the contributions of Ying Zhou and Colleen Kelly, Ph.D., Kelly Statistical Consulting in supporting the data analysis and the review of the manuscript. This study was funded by Sotera Wireless, Inc.,
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (
http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.