This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Medical Informatics, is properly cited. The complete bibliographic information, a link to the original publication on http://medinform.jmir.org/, as well as this copyright and license information must be included.
Although alarm safety is a critical issue that needs to be addressed to improve patient care, hospitals have not given serious consideration about how their staff should be using, setting, and responding to clinical alarms. Studies have indicated that 80%-99% of alarms in hospital units are false or clinically insignificant and do not represent real danger for patients, leading caregivers to miss relevant alarms that might indicate significant harmful events. The lack of use of any intelligent filter to detect recurrent, irrelevant, and/or false alarms before alerting health providers can culminate in a complex and overwhelming scenario of sensory overload for the medical team, known as
This paper’s main goal is to propose a solution to mitigate
We present a new approach to cope with the
The main contributions described are as follows: (1) a list of FAIs we defined that can be utilized and possibly extended by other researchers, (2) a novel approach to assess the probability of a false alarm using statistical analysis of multiple inputs representing alarm-context information, and (3) a reasoning algorithm that uses alarm-context information to detect false alarms in order to decide whether to notify caregivers with an indication of FAP (ie, FAP_LABEL) to avoid
Experiments were conducted to demonstrate that by providing an intelligent notification system, we could decide how to identify false alarms by analyzing alarm-context information. The reasoner entity we described in this paper was able to attribute FAP values to alarms based on FAIs and to notify caregivers with a FAP_LABEL indication without compromising patient safety.
In our previous work [
However, these alerts are often false alarms that do not represent real danger for patients. In this case, the lack of use of any intelligent filter to detect an indication of false alarms before alerting health providers can culminate in a context of a sensory overload for the medical team. This context can result in alarm fatigue and compromise health providers’ attention, leading them to miss relevant alarms that might announce significant harmful events.
As a strategy to mitigate the alarm fatigue issue, we present a new approach to monitor patients by using an intelligent notification process supported by a reasoning mechanism. This mechanism associates a false alarm probability (FAP) to alarms based on their real-time context information, including (1) information about a patient’s circumstances, such as his or her repositioning in bed, and localization, which is tracked in real time using wearable devices with GPS, and (2) information about sensors, including battery charge life, the last time the patient’s skin was prepared to receive electrodes, and the last time electrodes were changed, among others.
After receiving this context information as input, the reasoner’s work begins by analyzing each alarm and calculating the FAP associated with it according to the false alarm indicators (FAIs) we defined, based on our literature review. Thus, the reasoner uses the FAP calculated for each alarm to decide whether to include an indication of false alarm probability (ie, FAP_LABEL) with a notification that can be visualized by caregivers.
This paper’s main goal is to propose a solution to mitigate alarm fatigue by using an automatic reasoning mechanism to assist caregivers in their decision-making process of choosing which alarms they should respond to next. Our specific goal is to attribute an FAP to each alert based on the context in which it has been generated, such as a patient’s condition and information about monitoring devices and sensors. We aim to determine the probability of an alarm being a false alarm in order to decide whether to include this information (ie, FAP_LABEL) with the notifications sent to caregivers.
We addressed the following research questions: (1) How can an automatic reasoning system calculate an indication of FAP for an alarm generated by sensors and monitoring devices? (2) How can we decide whether to add an FAP_LABEL to a notification that could be visualized by the health care team?
We defined the following hypotheses for our case study:
Hypothesis 1 (H1): Our reasoning algorithm should associate an FAP value to every alarm generated by sensors and monitoring devices in our experiments.
Hypothesis 2 (H2): Our reasoning algorithm should add an indication of an FAP to each alarm, upon which the reasoner should decide whether or not to notify caregivers with an indication of FAP (ie, FAP_LABEL).
Hypothesis 3 (H3): Patient safety should not be compromised if and when the reasoning algorithm decides to add an FAP_LABEL to the notification.
The main contributions of this work are as follows:
A list of the FAIs we defined that can be utilized and possibly extended by other researchers.
A novel approach to assess the probability of a false alarm using statistical analysis of multiple inputs representing alarm-context information.
A reasoning algorithm that uses alarm-context information to detect false alarms in order to decide whether to notify caregivers with an indication of FAP (ie, FAP_LABEL) to avoid alarm fatigue.
Alarms are utilized to improve patient safety and quality of care by detecting changes early and requiring appropriate action. However, the medical literature contains many studies showing that up to 90% of all alarms in critical-care monitoring are false positives. The vast majority of all threshold alarms in the intensive care unit (ICU) do not have a real clinical impact on the care of the critically ill [
Many studies have recorded the number of alerts being triggered nowadays in ICUs during a period of time in order to analyze the impact of alarm safety in patient care as a consequence of the excessive volume of alarms. For instance, Kierra reported that during a 12-day analysis of the alarm system at The Johns Hopkins Hospital in Baltimore, USA, there was an average of 350 alerts per bed per day and that in one ICU, the average was 771 alerts per bed per day [
Lawless analyzed alarm soundings that occurred in an ICU during a 7-day period, recorded by ICU staff [
In addition to the excessive number of alarms in ICUs, another alarm-related problem, as presented by Sendelbach, is the high number of different alarm signals that was reducing the effectiveness of the alarms, creating confusion for staff, and was thus detrimental to patient care [
The main problem of having so many different devices triggering alarms is that it is not feasible for nurses to identify all of them, which means that this increase has occurred despite staff having difficulty in learning all available alarm signals in their work environment. Staff from an OR were only able to identify between 10 and 15 out of the 26 alarms triggered in the room, and ICU nurses could only identify between 9 and 14 out of 23 alarms found in the ICU, which contributes to the alarm overload problem [
Kerr and Hayes [
Lastly, a third alarm-related problem we are focusing on in this paper is the excessive number of false alarms. Studies have indicated that false and/or clinically insignificant alarms range from 80% to 99% [
Along with the already-mentioned alarm-related problems that can affect patient care, there is more information in ICUs that is considered critical for the health care team, such as (1) the perceived alarm urgency, and (2) the perceived true alarm rate of the alarm system [
Regarding the perceived true alarm rate of the alarm system, an important finding by Tanner is the link between the impact of the perceived true alarm rate of the alarm system by caregivers and its influence on patient care. The author showed that the nurses’ responses to alarms follow the perceived true alarm rate of the alarm system. According to the author, if the true alarm rate is perceived to be 10% reliable, then the response rate will be about 10% [
Although alarm safety is a critical issue that needs to be addressed to improve patient care, hospitals have not given serious consideration to how their staff should be using, setting, and responding to clinical alarms, according to the Emergency Care Research Institute (ECRI) [
By definition,
To illustrate this scenario, studies have indicated that false and/or clinically insignificant alarms range from 80% to 99% [
In
Summary of alarm-related issues.
Alarm-related issue | Causes | Consequences to the staff | Consequences to patient care | Avoidance strategies |
Excessive false positive alarms | Can be attributed to patient manipulation (ie, motion artifact) | Apathy and desensitization |
Reduction in responding |
Suspension of alarms for a short period prior to patient manipulation |
Frequent insignificant or irrelevant alarms | Use of the default alarm settings |
Distraction |
Disruption of patient care |
Eliminating nonessential alarms |
According to Imhoff et al, the quality of medical device alarms is unsatisfactory, affecting quality of care and patient safety. Since the low quality of alarm-generating algorithms is one of the main causes of this problem, major improvements in alarm algorithms are urgently needed [
To achieve this goal, a variety of alarm-suppression algorithms have been developed and successfully applied in the laboratory and the clinical environment to avoid alarm fatigue, such as relevance vector machine learning, statistical metrics, time series analysis, spectral regression, feature selection, and other classifiers [
Another strategy to avoid alarm fatigue is to use notification delays that are performed through the use of a middleware between the alarming medical device and the clinicians’ receiver device, such as a mobile phone or a tablet. Several studies found that introducing alarm delays before notifying caregivers could decrease
According to Imhoff et al, one of the main areas in which alarms can be improved is alarm validation (ie, determining whether the alarm is actually valid) [
With regard to methodology, we present a new approach to mitigate the alarm fatigue issue. We developed an application that attributes an FAP to alarms based on FAIs that we defined. Our reasoning algorithm uses the calculated FAP to decide whether to include an indication of FAP with a notification (ie, FAP_LABEL) before sending it to caregivers, in order to assist them in the complex task of choosing the next alarms to which they should respond.
In our system, a notification is a type of message that is sent to caregivers and contains information about a detected alarm or a group of alarms. An FAP is associated with an individual alarm; we calculate the FAP according to the FAIs we describe next, while an FAP_LABEL, on the other hand, corresponds to the probability of a notification containing a false alarm.
We calculate the FAP of every alarm triggered by our system. However, the reasoning algorithm decides whether to include the indication of the FAP with a notification—as the FAP_LABEL—based on the FAIs. The FAP_LABEL is the piece of information that can be visualized by caregivers. The inputs for our algorithm are a notification and its context information, including information about the patient’s conditions and sensors. After receiving these inputs, the reasoner starts working by analyzing the notification content and calculating the FAP_LABEL associated with it.
The processes to calculate the FAP and the FAP_LABEL are described below.
State machine diagram showing how we calculate the false alarm probability (FAP) associated with an alarm. FAI: false alarm indicator.
To calculate the FAP associated with each alarm, we defined four indicators of false alarms based on the information we gathered in our literature review. According to Kerr and Hayes, the main events that cause false alarms are patient movement or repositioning in bed and poor placement of sensors. Another common issue that triggers alarms is related to technical problems, such as lack of a battery in the monitoring devices [
The four FAIs defined in this case study represent information about (1) the duration of a sensor battery and the last time it was changed, (2) the last time the patient’s skin was prepared to receive electrodes and the last time they were changed, (3) the patient’s mobility, and (4) the patient’s position in bed. To calculate the FAI percentage in our experiments, we considered each indicator to have the same weight. The FAIs are listed below:
FAI1: Sensor battery FAI (SENSOR_BATTERY_FAI). This is an indication of the FAP associated with the battery-charge level of the sensors attached to the patient.
FAI2: Placement of sensor FAI (PLACEMENT_OF_SENSOR_FAI). FAI2 is related to the placement of a sensor (ie, whether a sensor is properly in touch with the patient’s skin).
FAI3: Patient mobility FAI (PATIENT_MOBILITY_FAI). This indicator is related to patient mobility, which means that it can evaluate the probability that the alarm has been triggered due to his or her movement from the bed to other places.
FAI4: Patient repositioning FAI (PATIENT_REPOSITIONING_FAI). This indicator can be used to calculate the FAP related to patient repositioning (ie, whether the alarm has been sent simply because the patient may have changed his or her position in bed).
As shown in
Inputs for our reasoning algorithm.
Input | Input name | FAIa the input is used to calculate | Description | Type of related monitoring device |
1 | LEVEL_OF_ BATTERY | FAI1 |
Level of battery for each monitoring device, including multiparametric monitors | Monitoring devices that use batteries |
2 | LAST_TIME_ BATTERY_ CHANGED | FAI1 |
Last time the device’s battery was changed | Monitoring devices that use batteries |
3 | LAST_TIME_ SKIN_ PREPARATION | FAI2 |
Last time skin preparation occurred | Sensors that use electrodes |
4 | LAST_TIME_ ELECTRODES_ CHANGED | FAI2 |
Last time electrodes were changed | Sensors that use electrodes |
5 | CURRENT_ PATIENT_ LOCALIZATION | FAI3 |
The current patient’s localization | Sensors used to track patient localization |
6 | LOG_LAST_ PATIENT_ LOCALIZATION | FAI3 |
A log of the patient’s last localization | Sensors used to track patient localization |
7 | CURRENT_ PATIENT_POSITION_IN_BED | FAI4 |
The current position a patient occupies in a bed | Sensors used to track patient position in bed |
8 | LOG_LAST_ PATIENT_POSITIONS_IN_BED | FAI4 |
The last positions a patient has occupied in a bed | Sensors used to track patient position in bed |
aFAI: false alarm indicator.
There is one output of our algorithm—Output1: The probability that an alarm is false (ie, the FAP).
To test our reasoning algorithm, we developed a system comprising an application (ie, the Producer App) that sends alarms to a broker who routes them to consumer applications that receive these alarms on behalf of the health care team. The system was developed in the Java language using the RabbitMQ message broker (Pivotal) [
The application scenario consisted of a group of four patients being monitored in an ICU with sensors and monitoring devices, such as multiparametric monitors (see
The Philips Efficia CM100 monitor.
Arduino UNO microcontroller.
The Philips Efficia CM100 monitor [
The electronic health (eHealth) Sensor Platform Complete Kit, version 2.0 (Cooking Hacks) [
eHealth Sensor Platform Complete Kit.
eHealth Sensor Shield.
In our simulated environment, patients were monitored through the use of two sensors: heart rate and temperature. The sensor readings were generated by a vital signs simulator that we developed. Regarding the sensor data simulated for each sensor, the temperature readings were generated randomly by the simulator within the 35.0°C-42.0°C range and the heart rate readings were randomly selected from the 40-188 beats-per-minute range. To define when a given temperature and heart rate reading represented an anomalous value that should trigger an alarm, we defined the thresholds shown in
Defining the anomaly thresholds of temperature and heart rate sensors for each patient.
Patient ID | Minimum temperature, °C | Maximum temperature, °C | Minimum heart rate, BPMa | Maximum heart rate, BPM |
1 | 35.5 | 39.0 | 60 | 100 |
2 | 35.0 | 38.5 | 55 | 95 |
3 | 35.5 | 39.5 | 60 | 100 |
4 | 35.5 | 38.5 | 50 | 100 |
aBPM: beats per minute.
In our experiments, we set the FAP_NOT_MIN at 75% (ie, the minimum value used as a reference to decide whether to add the FAP_LABEL to the notification). This means that every time the calculated FAP for an alarm was higher than or equal to 75%, our reasoner added the FAP_LABEL to the notification. Otherwise, we set the FAP_LABEL in our dataset to UNDEFINED, meaning that it was not included in the notification as an additional piece of information for caregivers (see
In
Results of our experiments for notifications related to temperature alarms.
NIDa | WIDb | PIDc | AIDd | Sensor value, °C | Alarm timestamp, date and time | FAPe,% | Notification timestamp, date and time | FAP_LABEL, % |
1 | 1 | 1 | 1 | 35.0 | 2019-07-02 |
50.0 | 2019-07-02 |
UNDEFINED |
2 | 1 | 4 | 2 | 42.0 | 2019-07-02 |
25.0 | 2019-07-02 |
UNDEFINED |
3 | 1 | 3 | 4 | 41.0 | 2019-07-02 |
50.0 | 2019-07-02 |
UNDEFINED |
4 | 1 | 2 | 9 | 41.0 | 2019-07-02 |
75.0 | 2019-07-02 |
75.0 |
5 | 1 | 1 | 12 | 42.0 | 2019-07-02 |
50.0 | 2019-07-02 |
UNDEFINED |
5 | 1 | 1 | 15 | 42.0 | 2019-07-02 |
100.0 | 2019-07-02 |
100.0 |
5 | 1 | 1 | 16 | 41.0 | 2019-07-02 |
75.0 | 2019-07-02 |
75.0 |
5 | 1 | 1 | 17 | 35.0 | 2019-07-02 |
25.0 | 2019-07-02 |
UNDEFINED |
5 | 1 | 1 | 18 | 42.0 | 2019-07-02 |
50.0 | 2019-07-02 |
UNDEFINED |
5 | 1 | 1 | 20 | 41.0 | 2019-07-02 |
50.0 | 2019-07-02 |
UNDEFINED |
aNID: Notification ID.
bWID: Ward ID.
cPID: Patient ID.
dAID: Alarm ID.
eFAP: false alarm probability.
Results of our experiments for notifications related to heart rate vital signs.
NIDa | WIDb | PIDc | AIDd | Sensor value, BPMe | Alarm timestamp, date and time | FAPf,% | Notification timestamp, date and time | FAP_LABEL, % |
1 | 1 | 2 | 1 | 108.0 | 2019-07-02 |
75.0 | 2019-07-02 |
75.0 |
2 | 1 | 1 | 2 | 145.0 | 2019-07-02 |
25.0 | 2019-07-02 |
UNDEFINED |
3 | 1 | 4 | 6 | 123.0 | 2019-07-02 |
50.0 | 2019-07-02 |
UNDEFINED |
4 | 1 | 3 | 8 | 116.0 | 2019-07-02 |
50.0 | 2019-07-02 |
UNDEFINED |
5 | 1 | 2 | 3 | 156.0 | 2019-07-02 |
0.0 | 2019-07-02 |
UNDEFINED |
5 | 1 | 2 | 5 | 159.0 | 2019-07-02 |
50.0 | 2019-07-02 |
UNDEFINED |
5 | 1 | 2 | 7 | 44.0 | 2019-07-02 |
75.0 | 2019-07-02 |
75.0 |
5 | 1 | 2 | 9 | 164.0 | 2019-07-02 |
50.0 | 2019-07-02 |
UNDEFINED |
5 | 1 | 2 | 16 | 184.0 | 2019-07-02 |
25.0 | 2019-07-02 |
UNDEFINED |
5 | 1 | 2 | 23 | 51.0 | 2019-07-02 |
0.0 | 2019-07-02 |
UNDEFINED |
aNID: Notification ID.
bWID: Ward ID.
cPID: Patient ID.
dAID: Alarm ID.
eBPM: beats per minute.
fFAP: false alarm probability.
Alarm safety is a complex problem to solve, influenced by a number of factors that extrapolate technology challenges and limitations, such as human influences, difficult patient conditions, a wide variety of environmental conditions, and even staffing cultures [
In this work, we tried to fill the gap of having feasible solutions to mitigate the alarm fatigue problem by focusing on the issue of false positive alarms, known to be a serious problem that yet remains unsolved. This paper presented a reasoning algorithm to detect false alarms based on alarm-context information provided automatically by the use of sensors and wearable devices and manually by the inputs of caregivers.
In our experiments, we created a database of simulated alarm-context information to establish a basis for the development of our algorithm in order to confirm H1 and H2 in experimental settings. As we can see in the FAP column of
Regarding H3, which declares that patient safety will not be compromised if and when the reasoning algorithm decides to add an FAP_LABEL to the notification, we can assume that is confirmed, since our algorithm does not stop an alarm from being triggered even when the calculated FAP is considered very high. We can see an example of this information in the sixth row of
As future work, we are planning to evolve our solution to support an optimized version of our reasoning algorithm that calculates the optimal FAP_NOT_MIN based on the real-time volume of alarms being triggered in an ICU.
Another plan for future work is to develop a machine learning–based algorithm capable of predicting both the FAP and FAP_LABEL based on a dataset that contains the ICU information history, such as patients’ conditions, sensors, and alarms.
artificial intelligence
Alarm ID
National Council of Research Development
electrocardiogram
Emergency Care Research Institute
electronic health
electromyography
end-tidal carbon dioxide
false alarm indicator
false alarm probability
Foundation for Research Support of the State of Rio de Janeiro
Hypothesis 1
Hypothesis 2
Hypothesis 3
invasive blood pressure
intensive care unit
noninvasive blood pressure
Notification ID
operating room
Patient ID
peripheral oxygen saturation
Ward ID
This work was supported by grants from two Brazilian government agencies: CNPq (National Council of Research Development) and FAPERJ (Foundation for Research Support of the State of Rio de Janeiro).
None declared.