Intensive care unit (ICU) alarms were designed to call attention to a patient, to alert a change in the patient's physiology or to alert staff to a device problem. Alarms are triggered when a physiologic variable crosses a set threshold. In their excellent literature review, Imhoff and Kuhls report alarm frequencies of 1.6 to 14.6 alarms/h and a false alarm rate of up to 90%.1 Chambrin et al.2 reported the lowest rate of alarms at 1.6 alarms/h, however, their study did not include infusion pumps (InfP) or alerts. Tsien and Fackler3 reported one of the highest alarm rates at 9.8 alarms/h in a noisier environment, but limited their study to alarms from the cardiac patient monitor. The problem with simple threshold alarms is that up to 94.5% of the alarms that sound in the ICU are false, are provider-induced,4 and frequently sound unnecessarily.1,2,4 Default settings by the equipment manufacturers are set to avoid missing a single false negative alarm and thereby result in many false positive alarms.5
New alarm algorithms and improvements in sensors are reported to reduce the number of false alarms, but many of these suggestions have not been incorporated into current monitors nor have their improvements been evaluated in patients.1 Rheineck-Leyssius and Kalkman6 proposed a highly effective method for reducing pulse oximeter (Spo2) alarms by introducing a 6-s delay thereby reducing alarm rates by 50%. One of the new and interesting approaches to reducing the number of false alarms is the use of context awareness.7,8 Dey8 defines context-awareness as: “A system is context-aware if it uses context to provide relevant information and/or services to the user, where relevancy depends on the user's task.” Chambrin et al.2 report that 42% of the transient ICU alarms are triggered by patient movement or respiratory effort. Therefore, an alarm system that knows the patient is moving or coughing could suppress many motion-induced alarms. Although other investigators2–4,9,10 have classified false alarms into general categories, such as “staff manipulation” or “the patient,” we propose using specific tasks performed by the health care provider and each patient's current condition and actions. Some work regarding alarms and their context has been performed. For example, Seagull and Sanderson11 investigated anesthesia alarms in the context of the surgical phase (induction, maintenance, emergence). However, there is still more to explore in the ICU setting.
The purpose of this study was to observe alarms in the medical ICU (MICU) to identify methods for reducing the number of false alarms by using time delays and the correlations between alarms and clinical context.
Approval was obtained from the University of Utah Health Sciences Center's IRB and informed consent was obtained from 22 participating health care team members.
At the beginning of each day, for 24 days, the investigator randomly selected a patient room in the MICU, where a tracheally intubated patient was receiving respiratory support. A different patient and room were chosen every morning, except one patient who was observed twice. The investigator recorded health care team members' actions while they were in the patient's room and whether they came into the room in response to an alarm. Health care team members included attending physicians, fellow physicians, resident physicians, nurses, respiratory therapists, health care assistants, physical therapists, medical students, pharmacists, and other providers. Observations began at approximately 7:30 am and ended before 7 pm.
The 12-bed adult MICU is organized in an H shape, with individual patient rooms to the north and south, a central station in its center, and additional function rooms between the two rows of rooms. The doors to the patient's room were left open unless procedures were performed or privacy was required. The unit was staffed with one nurse for every two patients, one health care assistant, and one health unit coordinator. Respiratory therapists checked a patient's ventilator when paged or at least once every 4 h. Most patients had sepsis, respiratory failure, acute respiratory distress syndrome, multisystem organ failure, or renal failure. Approximately 25% of the patients had myocardial infarction, cardiomyopathy, or arrhythmias.
A cardiac monitor with at least electrocardiography, Spo2, and noninvasive arterial blood pressure (NBP) modules was present in each patient's room (HP M1094B, Philips Medical Systems, N.A., Bothell, WA). The unit's central monitoring station was generally not staffed. Ventilators included a Siemens Servo 300/300A (Draeger Medical, Telford, PA), a Nellcor Puritan Bennett 840 (Nellcor Puritan Bennett LLC, Pleasanton, CA), or a Viasys Avea (VIASYS Healthcare, Conshohocken, PA). Alaris Medley infusion pumps were used in every room (Cardinal Health, Dublin, OH). Flexiflo Quantum feeding pumps (Abbott Laboratories, Abbott Park, IL) were used in 13 observed rooms.
Time-stamped detailed information of alarms and the presence of health care team members were recorded manually using a COMPAQ iPAQ Pocket PC (Hewlett-Packard Company, Palo Alto, CA) and abcDB Database v.6.0 (PocketSOFT.ca, Lloydminster, SA, Canada). For health care team members, the time of entrance and exit as well as the provider category were recorded using a predefined list. When an alarm occurred, the observer recorded the device sounding the alarm, the alarm threshold settings, the alarm cause if identifiable, and the variable that produced the alarm: heart rate, Spo2, arterial blood pressure or NBP, pulmonary artery pressure, central venous pressure, temperature, peak airway pressure, minute volume (MV), tidal volume (TV), respiratory rate (RR) and apnea, InfP faults and feeding pump (FeedP) faults. For bedside tasks, the observer selected interventions from a predefined list and added free text comments with more detail. The following task categories were used: device alarm silenced, drug administered/dosage changed, patient assessment, physical therapy, washing, oral care, patient monitor settings changed, ventilator settings changed, data charted, arterial blood gas drawn, blood glucose levels measured, patient repositioned, airway suctioned, or other action taken.
During the study, the observer classified each alarm as true, true irrelevant or false. However, the observer was not a clinician, so all alarms were reclassified after the conclusion of the study using the following categories: effective, ineffective, or ignored. An alarm was classified as effective when an alarm-related action was performed by a qualified health care provider within 5 min of the end of the alarm. A qualified provider is one who has the authority to take alarm-related action. For example, physical therapists, phlebotomists, and health care assistants were only qualified to call for assistance, whereas nurses were qualified to administer medications, suction the patient's airway and change patient monitor settings. Only respiratory therapists and physicians were qualified to change ventilator settings.
Effective alarms were separated into two categories based on the action performed: (a) Technical actions include restarting infusion pumps, changing alarm thresholds, re-measuring values, changing sensor positions, reconnecting breathing circuits and all other equipment-related actions, and (b) patient actions included giving sedatives to an agitated patient, suctioning the airway, changing vasoactive drug infusion rates, repositioning agitated patients, and all other patient-related actions. An alarm was classified as ineffective if the alarm sounded, but a qualified health care provider did not enter the room in response to the alarm or was not present during the alarm. An alarm was classified as ignored when a qualified health care provider was present in the patient's room and no alarm-related action was taken during or within 5 min of the end of the alarm or the alarm was silenced from the nursing station and no action occurred.
Analysis of the data was performed using MATLAB (The MathWorks, Natick, MA). The pocket PC generated ACCESS/EXCEL files (Microsoft Corporation, Redmond, WA) were parsed, events were categorized and alarm start and end times were paired with the times a person entered and left the room.
Twenty-two health care team members participated in the study and gave written consent: 13 nurses, 3 nursing student interns, 3 respiratory therapists, 1 health care assistant, and 2 attending physicians. Several others, including phlebotomists, technicians and residents, who participated in the study gave verbal consent. Two-hundred hours of data were collected from 22 patients over 24 days (13 males and 9 females, mean age 54.6 ± 18.5 yr with a range from 21 to 93 yr). One day's data were lost and during 1 day participating health care team members did not care for a patient that met the inclusion criteria. Observations were made for an average of 9.16 h per day (range, 6.25–10.5 h). Two patients' lungs were ventilated using a Viasys Avea ventilator, 10 patients using a Siemens Servo 300 or 300A ventilator and 10 patients using a Nellcor Puritan Bennett 840 ventilator. Respiratory therapists, and occasionally the attending physicians or fellow physicians, changed the ventilator alarm thresholds; nurses changed the cardiac monitor alarm thresholds. We observed 10 changes to the patient monitor's alarm settings (5 NBP, 1 Spo2 and 4 not recorded) and 23 changes to ventilator alarm settings (8 MV, 4 peak airway pressure, 4 TV, 1 RR, 1 multiple changes, and 5 not recorded).
During the 200 h of observation, 1214 alarms occurred (6.07 alarms per hour): Table 1 shows that 5.3% were effective and patient-related, 17.7% were effective and technically related, 36.3% were ineffective, and 40.7% were ignored. Figure 1 shows the number of alarms generated by each variable and the length each alarm was active. The median alarm length was 17 s (range, 1 s to 17.25 min): 45.1% lasted for <15 s, 74.4% for <30 s, and 89.4% for <60 s. Of all the alarms, 34.3% ended without any health care team member being present in the patient's room, thus they canceled themselves when the alarming condition cleared. Many more alarms cleared when no health care team member qualified to respond to this alarm was present. Only the feeding pump and the infusion pump always required user intervention for the alarm to stop.
Figure 2 shows the total number of alarms for each of the four alarm types. A 19-s alarm delay would reduce the number of ignored and ineffective alarms by 67.1%, whereas a 14-s alarm delay would reduce it by 51.3%. For the effective alarms, the median time between the end of the alarm and the timestamp for the solution was 20 s; 77 solutions were performed before the alarm had ended.
Table 2 shows that ventilator manufacturers have taken different approaches for MV, TV and RR alarms. The Servo 300/300A ventilator does not alarm with TV or RR but with MV (TV times RR). The Nellcor 840 and Avea have separate alarms for all three related variables. As a consequence, the Nellcor 840 and Avea produced 4.29 alarms/h and 5.43 alarms/h, respectively, whereas the Servo 300 produced only 1.03 alarms/h. However, while the percentages of ineffective and ignored alarms of the 3 ventilators (Servo 300, Nellcor 840 and Avea) were similar (83%, 84%, and 88%), our observation periods were not equal (94, 90, and 16 h); therefore a statistical comparison was only performed between the Servo 300 and the Nellcor 840 group.
Unnecessary Alarms Occurring During Patient Care
During or within 2 min after suctioning, washing, repositioning, and oral care, 152 ineffective and ignored ventilator alarms were recorded (Table 3). A 2-min time window was chosen because the alarm silence button disabled alarms for 2 min. The primary alarm reason for 57 ventilator alarms, coded during the observation, was coughing. Patient's spontaneous breathing efforts were the cause for 118 ventilator alarms. Because no one was present in the room when 43.1% of the alarms started, they were not caused by a health care team member's actions.
Health Care Provider Presence and Tasks
During the 200-h study period, 1271 separate entries by a health care team member to the room being observed were recorded; their average stay was 4.6 min (range, from 2 s to 80.5 min). As seen in Figure 3, nurses made 65.7% of all visits; the patient's primary nurse made 44.8% of the visits. Of all providers, 15.6% stayed <30 s and 70.8% <5 min. Nurses contributed to the longest duration of health care team members' stay in the patient's room (62.6%, primary nurse 37.7%).
During the 200-h study period 2344 tasks were performed (Fig. 4). On average, 11.7 tasks per hour were performed (range, 6.9–21.5 task/h), and most were done by the nursing staff. The most common tasks were nurses administering medications or changing infusion rates (2.3/h), silencing alarms (1.3/h), charting (1.1/h), and patient assessments (0.7/h).
Of the 1214 alarms that occurred during our 200-h observation period, only 23% were effective. If the alarm onset would be delayed for 19 s, two-thirds of the ignored and ineffective alarms could have been avoided (Fig. 2). Suctioning, washing, repositioning, and oral care caused 152 ineffective and ignored ventilator alarms, of which 33 were longer than 19 s. If the alarm systems had been contextually aware of patient care procedures and waited 19 s before sounding an alarm, the combined ineffective and ignored alarm rate could have been reduced from 934 (77%) to 274 (50%) and the total number of alarms reduced by 54%. This reduction in the number of alarms would be clinically relevant as alarm noise has a detrimental effect on patient's sleep and ICU outcome.12 Additionally, this reduction should reduce alarm fatigue, a problem commonly observed in ICUs.4,13
Comparison with the Literature
Our observation of 6.1 alarms/h is consistent with a literature review by Imhoff and Kuhls1 reporting 1.6 to 14.6 alarms/h. We did not classify alarms as false and true; however, 77% of our alarms were ineffective and ignored alarms, which is similar to the false alarm rate of 90% reported by Imhoff and Kuhls.1
Alarm Classification Method
Tsien and Fackler3 define true, true irrelevant, and false alarms as: “True Positive, Clinically Relevant was used to indicate the monitoring device sounded an alarm, the alarm was appropriate given the actual data value as compared with the set threshold value, and the patient's condition required prompt attention…. True Positive, Clinically Irrelevant was used to indicate the monitor sounded an alarm, the alarm was appropriate given the input data value as compared with the set threshold value, but the patient's condition had not changed in a way that required additional medical attention…. False Positive was used to indicate that the monitor sounded an alarm, but the alarm was inappropriate given the input data value…. The alarm was false because the reported value did not reflect the patient condition.”
Classifying alarms into effective, ineffective, and ignored alarms has three advantages over the traditional method using true, true irrelevant, and false alarms: (a) the alarm classification can be performed by a trained observer rather than an expert clinician, (b) the classification can be performed after the completion of the study, as long as tasks and providers' actions are recorded, and (c) the criteria using a time cutoff and a task requirement related to the alarm makes it more objective than a single clinician's decision. However, there are also three disadvantages associated with this approach: (a) our method departs from the current alarm study literature using true and false alarms,1 (b) an effective alarm might be misclassified as an ineffective or ignored alarm if the response takes longer than 5 min to initiate, and (c) the alarm records must include tasks and health care provider actions.
Introducing an Alarm Delay
A delay would improve alarm reliability at the expense of lengthening the response time. It seemed that the staff currently respond selectively to alarms or wait before responding. They went to the patient's room in response to only 9.1% of the alarms, yet of these alarms 69.4% were effective alarms. It would be better for the alarm system to automatically introduce a delay rather than relying on the busy clinician to keep track of alarm duration. This proposal is consistent with a pulse oximeter study in which a 6-s delay reduced the alarm rate by 50%.12 Waiting 19 s before sounding an alarm would have reduced the number of Spo2 alarms by 52%. Newer Spo2 monitors claim to have reduced the false alarm rate to 15% with only minor delays by using better signal processing techniques.14,15 However, to keep the patient safe, asystole and ventilator disconnect/apnea alarms should be exempt from this delay.
Reducing Ventilator Alarms
TV was the most frequently occurring alarm; MV was the second most frequent (Table 1). The TV signal from the ventilator is a noisy signal, especially in patients with spontaneous breathing efforts and with active airway protection reflexes (coughing). TV alarms frequently occurred after suctioning. Waiting 19 s to announce a low TV would have had very little consequence to the patients we observed, as they all had MV and blood oxygenation saturation alarms that sound before desaturation occurs. The 19-s delay would have reduced the number of TV alarms in our observation period by 18% and the number of MV alarms by 37%. Apart from the patient's spontaneous respiratory efforts and coughing after suctioning, the leading causes for ventilator alarms were the lack of adaptations of the alarm threshold when ventilation modes were changed.
Perhaps TV and RR alarms are not necessary and a MV alarm is sufficient. The Servo 300, with a MV alarm, produced only 1.03 alarms/h, whereas the Nellcor 840 and Avea, with RR, TV and MV alarms, produced 4.29 alarms/h and 5.43 alarms/h, respectively (Table 2). However, without tracking patient outcome, we cannot say which strategy is best.
An even more conservative approach is taken by Philips with their Intellivue Event Surveillance system in which both MV and Spo2 must cross alarm thresholds before an event is identified (Philips Medical Systems, Andover, MA). This approach has been well accepted in neonatal care units. Our proposed approach would significantly improve the reliability of ventilator alarms and may result in more timely attention when the patient is truly at risk.
Reducing InfP and FeedP Alarms
InfP alarms were the longest in duration. One possibility to explain this behavior is that most InfP alarms did not identify a critical event, when the patient was in danger, and therefore the staff tended to ignore them for longer periods of time. InfP alarms had a high effective alarm rate (83%) because an alarm, once triggered, does not stop until the technical problem is resolved. We observed the nursing staff intentionally entering a smaller infusate volume than was available, so that the InfP alarm reminded them when the pump was nearly empty. Such alarm tailoring indicates the need for alarm redesign.16 Here manufacturers could implement a lower priority reminder function to support this behavior. In general, InfP alarms signaled mechanical failures and empty infusates, rather than patient trouble, and should be used only in situations involving the delivery of a life-supporting drug.
The FeedP alarms had a high effective alarm rate (90.3%) because there was no alarm silence button and the technical problem had to be fixed before the alarm would stop.
Reducing Alarms Occurring During Patient Care
Nursing care seems to generate a significant number of alarms. During our observations, 57% of the alarms occurred when a health care team member was in the room. Considering that nurses were in the room for only 18.3 min an hour, a disproportionate number of alarms occurred while they were in the room.
If the 2-min alarm silence button had been activated before suctioning, 74 alarms would have been prevented and the ventilator alarm rate would decrease from 2.8 alarms/h to 2.5 alarms/h.
Washing caused 10 unnecessary alarms; repositioning caused 59. If repositioning the patient were automatically detected by a mattress detection system,17,18 the number of unnecessary ventilator alarms could have been reduced by approximately 10%.
Health Care Provider Presence and Tasks
Figure 4 shows that silencing alarms constitutes approximately 16% of a nurse's bedside tasks. A radio frequency identification tracking system19 that could identify when a nurse arrives in the patient room and automatically silence alarms could reduce workload. However, caution is needed when changing the way alarms function because the new function might lead to unintended consequences known as “automation surprises.”20 Noise pollution could be reduced if alarms were to sound or be visually signaled outside the patient's room when a provider was not present in the room.
The number of ignored and ineffective alarms in a MICU could decrease from 934 to 274 by introducing a 19-s alarm delay, and by automatically detecting suctioning, patient repositioning, oral care, and blood gas sampling. Hopefully, more reliable alarms will elicit a more timely response, reduce workload, reduce noise pollution, and potentially improve patient safety.
The authors would like to thank all participating health care team members at the University of Utah Hospital for their support.
1. Imhoff M, Kuhls S. Alarm algorithms in critical care monitoring. Anesth Analg 2006;102:1525–37
2. Chambrin M, Ravaux P, Calvelo-Aros D, Jaborska A, Chopin C, Boniface B. Multicentric study of monitoring alarms in the adult intensive care unit (ICU): a descriptive analysis. Intensive Care Med 1999;25:1360–6
3. Tsien C, Fackler J. Poor prognosis for existing monitors in the intensive care unit. Crit Care Med 1997;25:614–19
4. Lawless S. Crying wolf: false alarms in a pediatric intensive care unit. Crit Care Med 1994;22:981–5
5. Quinn M. Semipractical alarms: a parable. J Clin Monit 1989;5:196–200
6. Rheineck-Leyssius A, Kalkman C. Influence of pulse oximeter settings on the frequency of alarms and detection of hypoxemia: theoretical effects of artifact rejection, alarm delay, averaging, median filtering or a lower setting of the alarm limit. J Clin Monit Comput 1998;14:151–6
7. Bricon-Souf N, Newman C. Context awareness in health care: a review. Int J Med Inform 2007;76:2–12
8. Dey AK. Providing Architectural Support for Building Context-Aware Applications [PhD thesis]. Atlanta (GA): Georgia Tech; 2000; 188
9. O'Carroll T. Survey of alarms in an intensive therapy unit. Anaesthesia 1986;41:742–4
10. Koski E, Mäkivirta A, Sukuvaara T, Kari A. Frequency and reliability of alarms in the monitoring of cardiac postoperative patients. Int J Clin Monit Comput 1990;7:129–33
11. Seagull FJ, Sanderson PM. Anesthesia alarms in context: an observational study. Hum Factors 2001;43:66–78
12. Christensen M. Noise levels in a general intensive care unit: a descriptive study. Nurs Crit Care 2007;12:188–97
13. Meredith C, Edworthy J. Are there too many alarms in the intensive care unit? An overview of the problems. J Adv Nurs 1995;21:15–20
14. Petterson M, Begnoche V, Graybeal J. The effect of motion on pulse oximetry and its clinical significance. Anesth Analg 2007;105(6 Suppl):S78–S84
15. Kawagishi T, Kanaya N, Nakayama M, Kurosawa S, Namiki A. A comparison of the failure times of pulse oximeters during blood pressure cuff-induced hypoperfusion in volunteers. Anesth Analg 2004;99:793–6
16. Watson M, Sanderson P, Russell WJ. Tailoring reveals information requirements: the case of anaesthesia alarms. Interact Comput 2004;16:271–93
17. Miyamoto Y, Yonezawa Y, Maki H, Ogawa H, Hahn A, Caldwell W. A system for monitoring heart pulse, respiration and posture in bed. Biomed Sci Instrum 2002;38:135–8
18. Watanabe K, Watanabe T, Watanabe H, Ando H, Ishikawa T, Kobayashi K. Noninvasive measurement of heartbeat, respiration, snoring and body movements of a subject in bed via a pneumatic method. IEEE Trans Biomed Engl 2005;52:2100–7
19. Goldman JM, Stahl E J, Sandberg WS, Fairbrother PL, Atkins PL, Egan MT, Wiklund RA, Rattner DW. Automatic Identification of the “Start of Anesthesia Care” using Indoor Positioning System (IPS) Derived Spatial-Temporal Association. Proceedings of the Annual Meeting of the Society for Technology in Anesthesiology; Santa Ana Pueblo, NM; 2004
© 2009 International Anesthesia Research Society
20. Cook RI, Woods DD. Adapting to new technology in the operating room. Hum Factors 1996;38:593–613