Vinay M. Nadkarni, MD, MS

Hospital physiologic monitors can alert clinicians to early signs of physiologic deterioration, and thus have great potential to save lives. However, monitors generate frequent alarms,[1, 2, 3, 4, 5, 6, 7, 8] and most are not relevant to the patient's safety (over 90% of pediatric intensive care unit (PICU)[1, 2] and over 70% of adult intensive care alarms).[5, 6] In psychology experiments, humans rapidly learn to ignore or respond more slowly to alarms when exposed to high false‐alarm rates, exhibiting alarm fatigue.[9, 10] In 2013, The Joint Commission named alarm fatigue the most common contributing factor to alarm‐related sentinel events in hospitals.[11, 12]

Although alarm fatigue has been implicated as a major threat to patient safety, little empirical data support its existence in hospitals. In this study, we aimed to determine if there was an association between nurses' recent exposure to nonactionable physiologic monitor alarms and their response time to future alarms for the same patients. This exploratory work was designed to inform future research in this area, acknowledging that the sample size would be too small for multivariable modeling.

METHODS

Study Definitions

The alarm classification scheme is shown in Figure 1. Note that, for clarity, we have intentionally avoided using the terms true and false alarms because their interpretations vary across studies and can be misleading.

Figure 1

Potentially Critical Alarm

A potentially critical alarm is any alarm for a clinical condition for which a timely response is important to determine if the alarm requires intervention to save the patient's life. This is based on the alarm type alone, including alarms for life‐threatening arrhythmias such as asystole and ventricular tachycardia, as well as alarms for vital signs outside the set limits. Supporting Table 1 in the online version of this article lists the breakdown of alarm types that we defined a priori as potentially and not potentially critical.

Table 1.Characteristics of the 2,445 Alarms for Clinical Conditions

	PICU				Ward
Alarm type	No.	% of Total	% Valid	% Actionable	No.	% of Total	% Valid	% Actionable
NOTE: Abbreviations: N/A, not applicable; PICU, pediatric intensive care unit.
Oxygen saturation	197	19.4	82.7	38.6	590	41.2	24.4	1.9
Heart rate	194	19.1	95.4	1.0	266	18.6	87.2	0.0
Respiratory rate	229	22.6	80.8	13.5	316	22.1	48.1	1.0
Blood pressure	259	25.5	83.8	5.8	11	0.8	72.7	0.0
Critical arrhythmia	1	0.1	0.0	0.0	4	0.3	0.0	0.0
Noncritical arrhythmia	71	7.0	2.8	0.0	244	17.1	8.6	0.0
Central venous pressure	49	4.8	0.0	0.0	0	0.0	N/A	N/A
Exhaled carbon dioxide	14	1.4	92.9	50.0	0	0.0	N/A	N/A
Total	1014	100.0	75.6	12.9	1,431	100.0	38.9	1.0

Valid Alarm

A valid alarm is any alarm that correctly identifies the physiologic status of the patient. Validity was based on waveform quality, lead signal strength indicators, and artifact conditions, referencing each monitor's operator's manual.

Actionable Alarm

An actionable alarm is any valid alarm for a clinical condition that either: (1) leads to a clinical intervention; (2) leads to a consultation with another clinician at the bedside (and thus visible on camera); or (3) is a situation that should have led to intervention or consultation, but the alarm was unwitnessed or misinterpreted by the staff at the bedside.

Nonactionable Alarm

An unactionable alarm is any alarm that does not meet the actionable definition above, including invalid alarms such as those caused by motion artifact, equipment/technical alarms, and alarms that are valid but nonactionable (nuisance alarms).[13]

Response Time

The response time is the time elapsed from when the alarm fired at the bedside to when the nurse entered the room or peered through a window or door, measured in seconds.

RESULTS

We performed 40 sessions among 40 patients and 36 nurses over 210 hours. We performed 20 sessions in children with heart and/or lung failure in the PICU and 20 sessions in children on a general ward. Sessions took place on weekdays between 9:00 _am and 6:00 _pm. There were 3 occasions when we filmed 2 patients cared for by the same nurse at the same time.

Nurses were mostly female (94.4%) and had between 2 months and 28 years of experience (median, 4.8 years). Patients on the ward ranged from 5 days to 5.4 years old (median, 6 months). Patients in the PICU ranged from 5 months to 16 years old (median, 2.5 years). Among the PICU patients, 14 (70%) were receiving mechanical ventilation only, 3 (15%) were receiving vasopressors only, and 3 (15%) were receiving mechanical ventilation and vasopressors.

We observed 5070 alarms during the 40 sessions. We excluded 108 (2.1%) that occurred at the end of video recording sessions with the nurse absent from the room because the nurse's response could not be determined. Alarms per session ranged from 10 to 1430 (median, 75; interquartile range [IQR], 35138). We excluded the outlier PICU patient with 1430 alarms in 5 hours from the analysis to avoid the potential for biasing the results. Figure 2 depicts the data flow.

Figure 2

Following the 5 training sessions, research assistants independently reviewed and made preliminary assessments on 4674 alarms; these alarms were all secondarily reviewed by a physician. Using the physician reviewer as the gold standard, the research assistant's sensitivity (assess alarm as actionable when physician also assesses as actionable) was 96.8% and specificity (assess alarm as nonactionable when physician also assesses as nonactionable) was 96.9%. We had to review 54 of 4674 alarms (1.2%) with an additional physician or nurse investigator to achieve consensus.

Characteristics of the 2445 alarms for clinical conditions are shown in Table 1. Only 12.9% of alarms in heart‐ and/or lung‐failure patients in the PICU were actionable, and only 1.0% of alarms in medical patients on a general inpatient ward were actionable.

Response‐Time Association With Nonactionable Alarm Exposure

Next, we analyzed the association between response time to potentially critical alarms that occurred when the nurse was not in the patient's room and the number of nonactionable alarms occurring over the preceding 120‐minute window. This required excluding the alarms that occurred in the first 120 minutes of each session, leaving 647 alarms with eligible response times to evaluate the exposure between prior nonactionable alarm exposure and response time: 219 in the PICU and 428 on the ward. Kaplan‐Meier plots and tabulated response times demonstrated the incremental relationships between each nonactionable alarm exposure category in the observed data, with the effects most prominent as the Kaplan‐Meier plots diverged beyond the median (Figure 3 and Table 2). Excluding the extreme outlier patient had no effect on the results, because 1378 of the 1430 alarms occurred with the nurse present at the bedside, and only 2 of the remaining alarms were potentially critical.

Figure 3

Table 2.Association Between Nonactionable Alarm Exposure in Preceding 120 Minutes and Response Time to Potentially Critical Alarms Based on Observed Data and With Response Time Modeled Using Weibull Accelerated Failure‐Time Regression

	Observed Data					Accelerated Failure‐Time Model
	Number of Potentially Critical Alarms	Minutes Elapsed Until This Percentage of Alarms Was Responded to				Modeled Response Time, min	95% CI, min	P Value*
50% (Median)		75%	90%	95%
NOTE: Abbreviations: CI, confidence interval; PICU, pediatric intensive care unit. *The critical P value used as the cut point between significant and nonsignificant, accounting for multiple comparisons, is 0.0167.
PICU
029 nonactionable alarms	70	1.6	8.0	18.6	25.1	2.8	1.9‐3.8	Reference
3079 nonactionable alarms	122	6.3	17.8	22.5	26.0	5.3	4.06.7	0.001 (vs 029)
80+ nonactionable alarms	27	16.0	28.4	32.0	33.1	8.5	4.312.7	0.009 (vs 029), 0.15 (vs 3079)
Ward
029 nonactionable alarms	159	9.8	17.8	25.0	28.9	7.7	6.39.1	Reference
3079 nonactionable alarms	211	11.6	22.4	44.6	63.2	11.5	9.613.3	0.001 (vs 029)
80+ nonactionable alarms	58	8.3	57.6	63.8	69.5	15.6	11.020.1	0.001 (vs 029), 0.09 (vs 3079)

Accelerated failure‐time regressions revealed significant incremental increases in the modeled response time as the number of preceding nonactionable alarms increased in both the PICU and ward settings (Table 2).

DISCUSSION

Alarm fatigue has recently generated interest among nurses,[22] physicians,[23] regulatory bodies,[24] patient safety organizations,[25] and even attorneys,[26] despite a lack of prior evidence linking nonactionable alarm exposure to response time or other adverse patient‐relevant outcomes. This study's main findings were that (1) the vast majority of alarms were nonactionable, (2) response time to alarms occurring while the nurse was out of the room increased as the number of nonactionable alarms over the preceding 120 minutes increased. These findings may be explained by alarm fatigue.

Our results build upon the findings of other related studies. The nonactionable alarm proportions we found were similar to other pediatric studies, reporting greater than 90% nonactionable alarms.[1, 2] One other study has reported a relationship between alarm exposure and response time. In that study, Voepel‐Lewis and colleagues evaluated nurse responses to pulse oximetry desaturation alarms in adult orthopedic surgery patients using time‐stamp data from their monitor notification system.[27] They found that alarm response time was significantly longer for patients in the highest quartile of alarms compared to those in lower quartiles. Our study provides new data suggesting a relationship between nonactionable alarm exposure and nurse response time.

Our study has several limitations. First, as a preliminary study to investigate feasibility and possible association, the sample of patients and nurses was necessarily limited and did not permit adjustment for nurse‐ or patient‐level covariates. A multivariable analysis with a larger sample might provide insight into alternate explanations for these findings other than alarm fatigue, including measures of nurse workload and patient factors (such as age and illness severity). Additional factors that are not as easily measured can also contribute to the complex decision of when and how to respond to alarms.[28, 29] Second, nurses were aware that they were being video recorded as part of a study of nonactionable alarms, although they did not know the specific details of measurement. Although this lack of blinding might lead to a Hawthorne‐like effect, our positive results suggest that this effect, if present, did not fully obscure the association. Third, all sessions took place on weekdays during daytime hours, but effects of nonactionable alarms might vary by time and day. Finally, we suspect that when nurses experience critical alarms that require them to intervene and rescue a patient, their response times to that patient's alarms that occur later in their shift will be quicker due to a heightened concern for the alarm being actionable. We were unable to explore that relationship in this analysis because the number of critical alarms requiring intervention was very small. This is a topic of future study.

CONCLUSIONS

We identified an association between a nurse's prior exposure to nonactionable alarms and response time to future alarms. This finding is consistent with alarm fatigue, but requires further study to more clearly delineate other factors that might confound or modify that relationship.

Hospital physiologic monitors can alert clinicians to early signs of physiologic deterioration, and thus have great potential to save lives. However, monitors generate frequent alarms,[1, 2, 3, 4, 5, 6, 7, 8] and most are not relevant to the patient's safety (over 90% of pediatric intensive care unit (PICU)[1, 2] and over 70% of adult intensive care alarms).[5, 6] In psychology experiments, humans rapidly learn to ignore or respond more slowly to alarms when exposed to high false‐alarm rates, exhibiting alarm fatigue.[9, 10] In 2013, The Joint Commission named alarm fatigue the most common contributing factor to alarm‐related sentinel events in hospitals.[11, 12]

Although alarm fatigue has been implicated as a major threat to patient safety, little empirical data support its existence in hospitals. In this study, we aimed to determine if there was an association between nurses' recent exposure to nonactionable physiologic monitor alarms and their response time to future alarms for the same patients. This exploratory work was designed to inform future research in this area, acknowledging that the sample size would be too small for multivariable modeling.

METHODS

Study Definitions

The alarm classification scheme is shown in Figure 1. Note that, for clarity, we have intentionally avoided using the terms true and false alarms because their interpretations vary across studies and can be misleading.

Figure 1

Potentially Critical Alarm

A potentially critical alarm is any alarm for a clinical condition for which a timely response is important to determine if the alarm requires intervention to save the patient's life. This is based on the alarm type alone, including alarms for life‐threatening arrhythmias such as asystole and ventricular tachycardia, as well as alarms for vital signs outside the set limits. Supporting Table 1 in the online version of this article lists the breakdown of alarm types that we defined a priori as potentially and not potentially critical.

Table 1.Characteristics of the 2,445 Alarms for Clinical Conditions

	PICU				Ward
Alarm type	No.	% of Total	% Valid	% Actionable	No.	% of Total	% Valid	% Actionable
NOTE: Abbreviations: N/A, not applicable; PICU, pediatric intensive care unit.
Oxygen saturation	197	19.4	82.7	38.6	590	41.2	24.4	1.9
Heart rate	194	19.1	95.4	1.0	266	18.6	87.2	0.0
Respiratory rate	229	22.6	80.8	13.5	316	22.1	48.1	1.0
Blood pressure	259	25.5	83.8	5.8	11	0.8	72.7	0.0
Critical arrhythmia	1	0.1	0.0	0.0	4	0.3	0.0	0.0
Noncritical arrhythmia	71	7.0	2.8	0.0	244	17.1	8.6	0.0
Central venous pressure	49	4.8	0.0	0.0	0	0.0	N/A	N/A
Exhaled carbon dioxide	14	1.4	92.9	50.0	0	0.0	N/A	N/A
Total	1014	100.0	75.6	12.9	1,431	100.0	38.9	1.0

Valid Alarm

A valid alarm is any alarm that correctly identifies the physiologic status of the patient. Validity was based on waveform quality, lead signal strength indicators, and artifact conditions, referencing each monitor's operator's manual.

Actionable Alarm

An actionable alarm is any valid alarm for a clinical condition that either: (1) leads to a clinical intervention; (2) leads to a consultation with another clinician at the bedside (and thus visible on camera); or (3) is a situation that should have led to intervention or consultation, but the alarm was unwitnessed or misinterpreted by the staff at the bedside.

Nonactionable Alarm

An unactionable alarm is any alarm that does not meet the actionable definition above, including invalid alarms such as those caused by motion artifact, equipment/technical alarms, and alarms that are valid but nonactionable (nuisance alarms).[13]

Response Time

The response time is the time elapsed from when the alarm fired at the bedside to when the nurse entered the room or peered through a window or door, measured in seconds.

RESULTS

We performed 40 sessions among 40 patients and 36 nurses over 210 hours. We performed 20 sessions in children with heart and/or lung failure in the PICU and 20 sessions in children on a general ward. Sessions took place on weekdays between 9:00 _am and 6:00 _pm. There were 3 occasions when we filmed 2 patients cared for by the same nurse at the same time.

Nurses were mostly female (94.4%) and had between 2 months and 28 years of experience (median, 4.8 years). Patients on the ward ranged from 5 days to 5.4 years old (median, 6 months). Patients in the PICU ranged from 5 months to 16 years old (median, 2.5 years). Among the PICU patients, 14 (70%) were receiving mechanical ventilation only, 3 (15%) were receiving vasopressors only, and 3 (15%) were receiving mechanical ventilation and vasopressors.

We observed 5070 alarms during the 40 sessions. We excluded 108 (2.1%) that occurred at the end of video recording sessions with the nurse absent from the room because the nurse's response could not be determined. Alarms per session ranged from 10 to 1430 (median, 75; interquartile range [IQR], 35138). We excluded the outlier PICU patient with 1430 alarms in 5 hours from the analysis to avoid the potential for biasing the results. Figure 2 depicts the data flow.

Figure 2

Following the 5 training sessions, research assistants independently reviewed and made preliminary assessments on 4674 alarms; these alarms were all secondarily reviewed by a physician. Using the physician reviewer as the gold standard, the research assistant's sensitivity (assess alarm as actionable when physician also assesses as actionable) was 96.8% and specificity (assess alarm as nonactionable when physician also assesses as nonactionable) was 96.9%. We had to review 54 of 4674 alarms (1.2%) with an additional physician or nurse investigator to achieve consensus.

Characteristics of the 2445 alarms for clinical conditions are shown in Table 1. Only 12.9% of alarms in heart‐ and/or lung‐failure patients in the PICU were actionable, and only 1.0% of alarms in medical patients on a general inpatient ward were actionable.

Response‐Time Association With Nonactionable Alarm Exposure

Next, we analyzed the association between response time to potentially critical alarms that occurred when the nurse was not in the patient's room and the number of nonactionable alarms occurring over the preceding 120‐minute window. This required excluding the alarms that occurred in the first 120 minutes of each session, leaving 647 alarms with eligible response times to evaluate the exposure between prior nonactionable alarm exposure and response time: 219 in the PICU and 428 on the ward. Kaplan‐Meier plots and tabulated response times demonstrated the incremental relationships between each nonactionable alarm exposure category in the observed data, with the effects most prominent as the Kaplan‐Meier plots diverged beyond the median (Figure 3 and Table 2). Excluding the extreme outlier patient had no effect on the results, because 1378 of the 1430 alarms occurred with the nurse present at the bedside, and only 2 of the remaining alarms were potentially critical.

Figure 3

Table 2.Association Between Nonactionable Alarm Exposure in Preceding 120 Minutes and Response Time to Potentially Critical Alarms Based on Observed Data and With Response Time Modeled Using Weibull Accelerated Failure‐Time Regression

	Observed Data					Accelerated Failure‐Time Model
	Number of Potentially Critical Alarms	Minutes Elapsed Until This Percentage of Alarms Was Responded to				Modeled Response Time, min	95% CI, min	P Value*
50% (Median)		75%	90%	95%
NOTE: Abbreviations: CI, confidence interval; PICU, pediatric intensive care unit. *The critical P value used as the cut point between significant and nonsignificant, accounting for multiple comparisons, is 0.0167.
PICU
029 nonactionable alarms	70	1.6	8.0	18.6	25.1	2.8	1.9‐3.8	Reference
3079 nonactionable alarms	122	6.3	17.8	22.5	26.0	5.3	4.06.7	0.001 (vs 029)
80+ nonactionable alarms	27	16.0	28.4	32.0	33.1	8.5	4.312.7	0.009 (vs 029), 0.15 (vs 3079)
Ward
029 nonactionable alarms	159	9.8	17.8	25.0	28.9	7.7	6.39.1	Reference
3079 nonactionable alarms	211	11.6	22.4	44.6	63.2	11.5	9.613.3	0.001 (vs 029)
80+ nonactionable alarms	58	8.3	57.6	63.8	69.5	15.6	11.020.1	0.001 (vs 029), 0.09 (vs 3079)

Accelerated failure‐time regressions revealed significant incremental increases in the modeled response time as the number of preceding nonactionable alarms increased in both the PICU and ward settings (Table 2).

DISCUSSION

Alarm fatigue has recently generated interest among nurses,[22] physicians,[23] regulatory bodies,[24] patient safety organizations,[25] and even attorneys,[26] despite a lack of prior evidence linking nonactionable alarm exposure to response time or other adverse patient‐relevant outcomes. This study's main findings were that (1) the vast majority of alarms were nonactionable, (2) response time to alarms occurring while the nurse was out of the room increased as the number of nonactionable alarms over the preceding 120 minutes increased. These findings may be explained by alarm fatigue.

Our results build upon the findings of other related studies. The nonactionable alarm proportions we found were similar to other pediatric studies, reporting greater than 90% nonactionable alarms.[1, 2] One other study has reported a relationship between alarm exposure and response time. In that study, Voepel‐Lewis and colleagues evaluated nurse responses to pulse oximetry desaturation alarms in adult orthopedic surgery patients using time‐stamp data from their monitor notification system.[27] They found that alarm response time was significantly longer for patients in the highest quartile of alarms compared to those in lower quartiles. Our study provides new data suggesting a relationship between nonactionable alarm exposure and nurse response time.

Our study has several limitations. First, as a preliminary study to investigate feasibility and possible association, the sample of patients and nurses was necessarily limited and did not permit adjustment for nurse‐ or patient‐level covariates. A multivariable analysis with a larger sample might provide insight into alternate explanations for these findings other than alarm fatigue, including measures of nurse workload and patient factors (such as age and illness severity). Additional factors that are not as easily measured can also contribute to the complex decision of when and how to respond to alarms.[28, 29] Second, nurses were aware that they were being video recorded as part of a study of nonactionable alarms, although they did not know the specific details of measurement. Although this lack of blinding might lead to a Hawthorne‐like effect, our positive results suggest that this effect, if present, did not fully obscure the association. Third, all sessions took place on weekdays during daytime hours, but effects of nonactionable alarms might vary by time and day. Finally, we suspect that when nurses experience critical alarms that require them to intervene and rescue a patient, their response times to that patient's alarms that occur later in their shift will be quicker due to a heightened concern for the alarm being actionable. We were unable to explore that relationship in this analysis because the number of critical alarms requiring intervention was very small. This is a topic of future study.

CONCLUSIONS

We identified an association between a nurse's prior exposure to nonactionable alarms and response time to future alarms. This finding is consistent with alarm fatigue, but requires further study to more clearly delineate other factors that might confound or modify that relationship.

Hospital physiologic monitors can alert clinicians to early signs of physiologic deterioration, and thus have great potential to save lives. However, monitors generate frequent alarms,[1, 2, 3, 4, 5, 6, 7, 8] and most are not relevant to the patient's safety (over 90% of pediatric intensive care unit (PICU)[1, 2] and over 70% of adult intensive care alarms).[5, 6] In psychology experiments, humans rapidly learn to ignore or respond more slowly to alarms when exposed to high false‐alarm rates, exhibiting alarm fatigue.[9, 10] In 2013, The Joint Commission named alarm fatigue the most common contributing factor to alarm‐related sentinel events in hospitals.[11, 12]

Although alarm fatigue has been implicated as a major threat to patient safety, little empirical data support its existence in hospitals. In this study, we aimed to determine if there was an association between nurses' recent exposure to nonactionable physiologic monitor alarms and their response time to future alarms for the same patients. This exploratory work was designed to inform future research in this area, acknowledging that the sample size would be too small for multivariable modeling.

METHODS

Study Definitions

The alarm classification scheme is shown in Figure 1. Note that, for clarity, we have intentionally avoided using the terms true and false alarms because their interpretations vary across studies and can be misleading.

Figure 1

Potentially Critical Alarm

A potentially critical alarm is any alarm for a clinical condition for which a timely response is important to determine if the alarm requires intervention to save the patient's life. This is based on the alarm type alone, including alarms for life‐threatening arrhythmias such as asystole and ventricular tachycardia, as well as alarms for vital signs outside the set limits. Supporting Table 1 in the online version of this article lists the breakdown of alarm types that we defined a priori as potentially and not potentially critical.

Table 1.Characteristics of the 2,445 Alarms for Clinical Conditions

	PICU				Ward
Alarm type	No.	% of Total	% Valid	% Actionable	No.	% of Total	% Valid	% Actionable
NOTE: Abbreviations: N/A, not applicable; PICU, pediatric intensive care unit.
Oxygen saturation	197	19.4	82.7	38.6	590	41.2	24.4	1.9
Heart rate	194	19.1	95.4	1.0	266	18.6	87.2	0.0
Respiratory rate	229	22.6	80.8	13.5	316	22.1	48.1	1.0
Blood pressure	259	25.5	83.8	5.8	11	0.8	72.7	0.0
Critical arrhythmia	1	0.1	0.0	0.0	4	0.3	0.0	0.0
Noncritical arrhythmia	71	7.0	2.8	0.0	244	17.1	8.6	0.0
Central venous pressure	49	4.8	0.0	0.0	0	0.0	N/A	N/A
Exhaled carbon dioxide	14	1.4	92.9	50.0	0	0.0	N/A	N/A
Total	1014	100.0	75.6	12.9	1,431	100.0	38.9	1.0

Valid Alarm

A valid alarm is any alarm that correctly identifies the physiologic status of the patient. Validity was based on waveform quality, lead signal strength indicators, and artifact conditions, referencing each monitor's operator's manual.

Actionable Alarm

An actionable alarm is any valid alarm for a clinical condition that either: (1) leads to a clinical intervention; (2) leads to a consultation with another clinician at the bedside (and thus visible on camera); or (3) is a situation that should have led to intervention or consultation, but the alarm was unwitnessed or misinterpreted by the staff at the bedside.

Nonactionable Alarm

An unactionable alarm is any alarm that does not meet the actionable definition above, including invalid alarms such as those caused by motion artifact, equipment/technical alarms, and alarms that are valid but nonactionable (nuisance alarms).[13]

Response Time

The response time is the time elapsed from when the alarm fired at the bedside to when the nurse entered the room or peered through a window or door, measured in seconds.

RESULTS

We performed 40 sessions among 40 patients and 36 nurses over 210 hours. We performed 20 sessions in children with heart and/or lung failure in the PICU and 20 sessions in children on a general ward. Sessions took place on weekdays between 9:00 _am and 6:00 _pm. There were 3 occasions when we filmed 2 patients cared for by the same nurse at the same time.

Nurses were mostly female (94.4%) and had between 2 months and 28 years of experience (median, 4.8 years). Patients on the ward ranged from 5 days to 5.4 years old (median, 6 months). Patients in the PICU ranged from 5 months to 16 years old (median, 2.5 years). Among the PICU patients, 14 (70%) were receiving mechanical ventilation only, 3 (15%) were receiving vasopressors only, and 3 (15%) were receiving mechanical ventilation and vasopressors.

We observed 5070 alarms during the 40 sessions. We excluded 108 (2.1%) that occurred at the end of video recording sessions with the nurse absent from the room because the nurse's response could not be determined. Alarms per session ranged from 10 to 1430 (median, 75; interquartile range [IQR], 35138). We excluded the outlier PICU patient with 1430 alarms in 5 hours from the analysis to avoid the potential for biasing the results. Figure 2 depicts the data flow.

Figure 2

Following the 5 training sessions, research assistants independently reviewed and made preliminary assessments on 4674 alarms; these alarms were all secondarily reviewed by a physician. Using the physician reviewer as the gold standard, the research assistant's sensitivity (assess alarm as actionable when physician also assesses as actionable) was 96.8% and specificity (assess alarm as nonactionable when physician also assesses as nonactionable) was 96.9%. We had to review 54 of 4674 alarms (1.2%) with an additional physician or nurse investigator to achieve consensus.

Characteristics of the 2445 alarms for clinical conditions are shown in Table 1. Only 12.9% of alarms in heart‐ and/or lung‐failure patients in the PICU were actionable, and only 1.0% of alarms in medical patients on a general inpatient ward were actionable.

Response‐Time Association With Nonactionable Alarm Exposure

Next, we analyzed the association between response time to potentially critical alarms that occurred when the nurse was not in the patient's room and the number of nonactionable alarms occurring over the preceding 120‐minute window. This required excluding the alarms that occurred in the first 120 minutes of each session, leaving 647 alarms with eligible response times to evaluate the exposure between prior nonactionable alarm exposure and response time: 219 in the PICU and 428 on the ward. Kaplan‐Meier plots and tabulated response times demonstrated the incremental relationships between each nonactionable alarm exposure category in the observed data, with the effects most prominent as the Kaplan‐Meier plots diverged beyond the median (Figure 3 and Table 2). Excluding the extreme outlier patient had no effect on the results, because 1378 of the 1430 alarms occurred with the nurse present at the bedside, and only 2 of the remaining alarms were potentially critical.

Figure 3

Table 2.Association Between Nonactionable Alarm Exposure in Preceding 120 Minutes and Response Time to Potentially Critical Alarms Based on Observed Data and With Response Time Modeled Using Weibull Accelerated Failure‐Time Regression

	Observed Data					Accelerated Failure‐Time Model
	Number of Potentially Critical Alarms	Minutes Elapsed Until This Percentage of Alarms Was Responded to				Modeled Response Time, min	95% CI, min	P Value*
50% (Median)		75%	90%	95%
NOTE: Abbreviations: CI, confidence interval; PICU, pediatric intensive care unit. *The critical P value used as the cut point between significant and nonsignificant, accounting for multiple comparisons, is 0.0167.
PICU
029 nonactionable alarms	70	1.6	8.0	18.6	25.1	2.8	1.9‐3.8	Reference
3079 nonactionable alarms	122	6.3	17.8	22.5	26.0	5.3	4.06.7	0.001 (vs 029)
80+ nonactionable alarms	27	16.0	28.4	32.0	33.1	8.5	4.312.7	0.009 (vs 029), 0.15 (vs 3079)
Ward
029 nonactionable alarms	159	9.8	17.8	25.0	28.9	7.7	6.39.1	Reference
3079 nonactionable alarms	211	11.6	22.4	44.6	63.2	11.5	9.613.3	0.001 (vs 029)
80+ nonactionable alarms	58	8.3	57.6	63.8	69.5	15.6	11.020.1	0.001 (vs 029), 0.09 (vs 3079)

Accelerated failure‐time regressions revealed significant incremental increases in the modeled response time as the number of preceding nonactionable alarms increased in both the PICU and ward settings (Table 2).

DISCUSSION

Alarm fatigue has recently generated interest among nurses,[22] physicians,[23] regulatory bodies,[24] patient safety organizations,[25] and even attorneys,[26] despite a lack of prior evidence linking nonactionable alarm exposure to response time or other adverse patient‐relevant outcomes. This study's main findings were that (1) the vast majority of alarms were nonactionable, (2) response time to alarms occurring while the nurse was out of the room increased as the number of nonactionable alarms over the preceding 120 minutes increased. These findings may be explained by alarm fatigue.

Our results build upon the findings of other related studies. The nonactionable alarm proportions we found were similar to other pediatric studies, reporting greater than 90% nonactionable alarms.[1, 2] One other study has reported a relationship between alarm exposure and response time. In that study, Voepel‐Lewis and colleagues evaluated nurse responses to pulse oximetry desaturation alarms in adult orthopedic surgery patients using time‐stamp data from their monitor notification system.[27] They found that alarm response time was significantly longer for patients in the highest quartile of alarms compared to those in lower quartiles. Our study provides new data suggesting a relationship between nonactionable alarm exposure and nurse response time.

Our study has several limitations. First, as a preliminary study to investigate feasibility and possible association, the sample of patients and nurses was necessarily limited and did not permit adjustment for nurse‐ or patient‐level covariates. A multivariable analysis with a larger sample might provide insight into alternate explanations for these findings other than alarm fatigue, including measures of nurse workload and patient factors (such as age and illness severity). Additional factors that are not as easily measured can also contribute to the complex decision of when and how to respond to alarms.[28, 29] Second, nurses were aware that they were being video recorded as part of a study of nonactionable alarms, although they did not know the specific details of measurement. Although this lack of blinding might lead to a Hawthorne‐like effect, our positive results suggest that this effect, if present, did not fully obscure the association. Third, all sessions took place on weekdays during daytime hours, but effects of nonactionable alarms might vary by time and day. Finally, we suspect that when nurses experience critical alarms that require them to intervene and rescue a patient, their response times to that patient's alarms that occur later in their shift will be quicker due to a heightened concern for the alarm being actionable. We were unable to explore that relationship in this analysis because the number of critical alarms requiring intervention was very small. This is a topic of future study.

CONCLUSIONS

We identified an association between a nurse's prior exposure to nonactionable alarms and response time to future alarms. This finding is consistent with alarm fatigue, but requires further study to more clearly delineate other factors that might confound or modify that relationship.

In recent years, many hospitals have implemented rapid response systems (RRSs) in efforts to reduce mortality outside the intensive care unit (ICU). Rapid response systems include 2 clinical components (efferent and afferent limbs) and 2 organizational components (process improvement and administrative limbs).[1, 2] The efferent limb includes medical emergency teams (METs) that can be summoned to hospital wards to rescue deteriorating patients. The afferent limb identifies patients at risk of deterioration using tools such as early warning scores and triggers a MET response when appropriate.[2] The process‐improvement limb evaluates and optimizes the RRS. The administrative limb implements the RRS and supports its ongoing operation. The effectiveness of most RRSs depends upon the ward team making the decision to escalate care by activating the MET. Barriers to activating the MET may include reduced situational awareness,[3, 4] hierarchical barriers to calling for help,[3, 4, 5, 6, 7, 8] fear of criticism,[3, 8, 9] and other hospital safety cultural barriers.[3, 4, 8]

Proactive critical‐care outreach[10, 11, 12, 13] or rover[14] teams seek to reduce barriers to activation and improve outcomes by systematically identifying and evaluating at‐risk patients without relying on requests for assistance from the ward team. Structured similarly to early warning scores, surveillance tools intended for rover teams might improve their ability to rapidly identify at‐risk patients throughout a hospital. They could combine vital signs with other variables, such as diagnostic and therapeutic interventions that reflect the ward team's early, evolving concern. In particular, the incorporation of medications associated with deterioration may enhance the performance of surveillance tools.

Medications may be associated with deterioration in one of several ways. They could play a causal role in deterioration (ie, opioids causing respiratory insufficiency), represent clinical worsening and anticipation of possible deterioration (ie, broad‐spectrum antibiotics for a positive blood culture), or represent rescue therapies for early deterioration (ie, antihistamines for allergic reactions). In each case, the associated therapeutic classes could be considered sentinel markers of clinical deterioration.

Combined with vital signs and other risk factors, therapeutic classes could serve as useful components of surveillance tools to detect signs of early, evolving deterioration and flag at‐risk patients for evaluation. As a first step, we sought to identify therapeutic classes associated with clinical deterioration. This effort to improve existing afferent tools falls within the process‐improvement limb of RRSs.

PATIENTS AND METHODS

Study Design

We performed a case‐crossover study of children who experienced clinical deterioration. An alternative to the matched case‐control design, the case‐crossover design involves longitudinal within‐subject comparisons exclusively of case subjects such that an individual serves as his or her own control. It is most effective when studying intermittent exposures that result in transient changes in the risk of an acute event,[15, 16, 17] making it appropriate for our study.

Using the case‐crossover design, we compared a discrete time period in close proximity to the deterioration event, called the hazard interval, with earlier time periods in the hospitalization, called the control intervals.[15, 16, 17] In our primary analysis (Figure 1B), we defined the durations of these intervals as follows: We first censored the 2 hours immediately preceding the clinical deterioration event (hours 0 to 2). We made this decision a priori to exclude medications used after deterioration was recognized and resuscitation had already begun. The 12‐hour period immediately preceding the censored interval was the hazard interval (hours 2 to 14). Each 12‐hour period immediately preceding the hazard interval was a control interval (hours 14 to 26, 26 to 38, 38 to 50, and 50 to 62). Depending on the child's length of stay prior to the deterioration event, each hazard interval had 14 control intervals for comparison. In sensitivity analysis, we altered the durations of these intervals (see below).

Figure 1

RESULTS

We identified 12 CPAs, 41 ARCs, and 699 ICU transfers during the study period. Of these 752 events, 141 (19%) were eligible as cases according to our inclusion criteria.[18] (A flowchart demonstrating the identification of eligible cases is provided in Supporting Table 25 in the online version of this article.) Of the 81% excluded, 37% were ICU transfers who did not meet urgent criteria. Another 31% were excluded because they were hospitalized for <24 hours at the time of the event, making their analysis in a case‐crossover design using 12‐hour periods impossible. Event characteristics, demographics, and hospitalization characteristics are shown in Table 2.

Sensitivity Analysis

Of the 14 classes that were significant in primary analysis, we carried 9 forward to sensitivity analysis. The 5 that were not carried forward overlapped substantially with other classes that were carried forward. The decision of which overlapping class to carry forward was based upon (1) parsimony and (2) clinical relevance. This is described briefly in the footnotes to Table 3 (see Supporting information in the online version of this article for a full description of this process). Figure 2 presents the odds ratios and their 95% confidence intervals for the sensitivity analysis of each therapeutic class that was significant in primary analysis. Loop diuretics remained significantly associated with deterioration in all 7 iterations. Glycopeptide antibiotics (vancomycin), third‐generation and fourth‐generation cephalosporins, systemic corticosteroids, and benzodiazepines were significant in 6. Anaerobic antibiotics and narcotic analgesics (full opioid agonists) were significant in 5, and aminoglycoside antibiotics and antidotes to hypersensitivity reactions (diphenhydramine) in 4.

Figure 2

DISCUSSION

We identified 9 therapeutic classes which were associated with a 2.5‐fold to 5.8‐fold increased risk of clinical deterioration. The results were robust to sensitivity analysis. Given their temporal association to the deterioration events, these therapeutic classes may serve as sentinels of early deterioration and are candidate variables to combine with vital signs and other risk factors in a surveillance tool for rover teams or an early warning score.

Although most early warning scores intended for use at the bedside are based upon vital signs and clinical observations, a few also include medications. Monaghan's Pediatric Early Warning Score, the basis for many modified scores used in children's hospitals throughout the world, assigns points for children requiring frequent doses of nebulized medication.[20, 21, 22] Nebulized epinephrine is a component of the Bristol Paediatric Early Warning Tool.[23] The number of medications administered in the preceding 24 hours was included in an early version of the Bedside Paediatric Early Warning System Score.[24] Adding IV antibiotics to the Maximum Modified Early Warning Score improved prediction of the need for higher care utilization among hospitalized adults.[25]

In order to determine the role of the IV medications we found to be associated with clinical deterioration, the necessary next step is to develop a multivariable predictive model to determine if they improve the performance of existing early warning scores in identifying deteriorating patients. Although simplicity is an important characteristic of hand‐calculated early warning scores, integration of a more complex scoring system with more variables, such as these medications, into the electronic health record would allow for automated scoring, eliminating the need to sacrifice score performance to keep the tool simple. Integration into the electronic health record would have the additional benefit of making the score available to clinicians who are not at the bedside. Such tools would be especially useful for remote surveillance for deterioration by critical‐care outreach or rover teams.

Our study has several limitations. First, the sample size was small, and although we sought to minimize the likelihood of chance associations by performing sensitivity analysis, these findings should be confirmed in a larger study. Second, we only evaluated IV medications. Medications administered by other routes could also be associated with clinical deterioration and should be analyzed in future studies. Third, we excluded children hospitalized for <24 hours, as well as transfers that did not meet urgent criteria. These may be limitations because (1) the first 24 hours of hospitalization may be a high‐risk period, and (2) patients who were on trajectories toward severe deterioration and received interventions that prevented further deterioration but did not meet urgent transfer criteria were excluded. It may be that the children we included as cases were at increased risk of deterioration that is either more difficult to recognize early, or more difficult to treat effectively without ICU interventions. Finally, we acknowledge that in some cases the therapeutic classes were associated with deterioration in a causal fashion, and in others the medications administered did not cause deterioration but were signs of therapeutic interventions that were initiated in response to clinical worsening. Identifying the specific indications for administration of drugs used in response to clinical worsening may have resulted in stronger associations with deterioration. However, these indications are often complex, multifactorial, and poorly documented in real time. This limits the ability to automate their detection using the electronic health record, the ultimate goal of this line of research.

CONCLUSION

We used a case‐crossover approach to identify therapeutic classes that are associated with increased risk of clinical deterioration in hospitalized children on pediatric wards. These sentinel therapeutic classes may serve as useful components of electronic health recordbased surveillance tools to detect signs of early, evolving deterioration and flag at‐risk patients for critical‐care outreach or rover team review. Future research should focus on evaluating whether including these therapeutic classes in early warning scores improves their accuracy in detecting signs of deterioration and determining if providing this information as clinical decision support improves patient outcomes.

In recent years, many hospitals have implemented rapid response systems (RRSs) in efforts to reduce mortality outside the intensive care unit (ICU). Rapid response systems include 2 clinical components (efferent and afferent limbs) and 2 organizational components (process improvement and administrative limbs).[1, 2] The efferent limb includes medical emergency teams (METs) that can be summoned to hospital wards to rescue deteriorating patients. The afferent limb identifies patients at risk of deterioration using tools such as early warning scores and triggers a MET response when appropriate.[2] The process‐improvement limb evaluates and optimizes the RRS. The administrative limb implements the RRS and supports its ongoing operation. The effectiveness of most RRSs depends upon the ward team making the decision to escalate care by activating the MET. Barriers to activating the MET may include reduced situational awareness,[3, 4] hierarchical barriers to calling for help,[3, 4, 5, 6, 7, 8] fear of criticism,[3, 8, 9] and other hospital safety cultural barriers.[3, 4, 8]

Proactive critical‐care outreach[10, 11, 12, 13] or rover[14] teams seek to reduce barriers to activation and improve outcomes by systematically identifying and evaluating at‐risk patients without relying on requests for assistance from the ward team. Structured similarly to early warning scores, surveillance tools intended for rover teams might improve their ability to rapidly identify at‐risk patients throughout a hospital. They could combine vital signs with other variables, such as diagnostic and therapeutic interventions that reflect the ward team's early, evolving concern. In particular, the incorporation of medications associated with deterioration may enhance the performance of surveillance tools.

Medications may be associated with deterioration in one of several ways. They could play a causal role in deterioration (ie, opioids causing respiratory insufficiency), represent clinical worsening and anticipation of possible deterioration (ie, broad‐spectrum antibiotics for a positive blood culture), or represent rescue therapies for early deterioration (ie, antihistamines for allergic reactions). In each case, the associated therapeutic classes could be considered sentinel markers of clinical deterioration.

Combined with vital signs and other risk factors, therapeutic classes could serve as useful components of surveillance tools to detect signs of early, evolving deterioration and flag at‐risk patients for evaluation. As a first step, we sought to identify therapeutic classes associated with clinical deterioration. This effort to improve existing afferent tools falls within the process‐improvement limb of RRSs.

PATIENTS AND METHODS

Study Design

We performed a case‐crossover study of children who experienced clinical deterioration. An alternative to the matched case‐control design, the case‐crossover design involves longitudinal within‐subject comparisons exclusively of case subjects such that an individual serves as his or her own control. It is most effective when studying intermittent exposures that result in transient changes in the risk of an acute event,[15, 16, 17] making it appropriate for our study.

Using the case‐crossover design, we compared a discrete time period in close proximity to the deterioration event, called the hazard interval, with earlier time periods in the hospitalization, called the control intervals.[15, 16, 17] In our primary analysis (Figure 1B), we defined the durations of these intervals as follows: We first censored the 2 hours immediately preceding the clinical deterioration event (hours 0 to 2). We made this decision a priori to exclude medications used after deterioration was recognized and resuscitation had already begun. The 12‐hour period immediately preceding the censored interval was the hazard interval (hours 2 to 14). Each 12‐hour period immediately preceding the hazard interval was a control interval (hours 14 to 26, 26 to 38, 38 to 50, and 50 to 62). Depending on the child's length of stay prior to the deterioration event, each hazard interval had 14 control intervals for comparison. In sensitivity analysis, we altered the durations of these intervals (see below).

Figure 1

RESULTS

We identified 12 CPAs, 41 ARCs, and 699 ICU transfers during the study period. Of these 752 events, 141 (19%) were eligible as cases according to our inclusion criteria.[18] (A flowchart demonstrating the identification of eligible cases is provided in Supporting Table 25 in the online version of this article.) Of the 81% excluded, 37% were ICU transfers who did not meet urgent criteria. Another 31% were excluded because they were hospitalized for <24 hours at the time of the event, making their analysis in a case‐crossover design using 12‐hour periods impossible. Event characteristics, demographics, and hospitalization characteristics are shown in Table 2.

Sensitivity Analysis

Of the 14 classes that were significant in primary analysis, we carried 9 forward to sensitivity analysis. The 5 that were not carried forward overlapped substantially with other classes that were carried forward. The decision of which overlapping class to carry forward was based upon (1) parsimony and (2) clinical relevance. This is described briefly in the footnotes to Table 3 (see Supporting information in the online version of this article for a full description of this process). Figure 2 presents the odds ratios and their 95% confidence intervals for the sensitivity analysis of each therapeutic class that was significant in primary analysis. Loop diuretics remained significantly associated with deterioration in all 7 iterations. Glycopeptide antibiotics (vancomycin), third‐generation and fourth‐generation cephalosporins, systemic corticosteroids, and benzodiazepines were significant in 6. Anaerobic antibiotics and narcotic analgesics (full opioid agonists) were significant in 5, and aminoglycoside antibiotics and antidotes to hypersensitivity reactions (diphenhydramine) in 4.

Figure 2

DISCUSSION

We identified 9 therapeutic classes which were associated with a 2.5‐fold to 5.8‐fold increased risk of clinical deterioration. The results were robust to sensitivity analysis. Given their temporal association to the deterioration events, these therapeutic classes may serve as sentinels of early deterioration and are candidate variables to combine with vital signs and other risk factors in a surveillance tool for rover teams or an early warning score.

Although most early warning scores intended for use at the bedside are based upon vital signs and clinical observations, a few also include medications. Monaghan's Pediatric Early Warning Score, the basis for many modified scores used in children's hospitals throughout the world, assigns points for children requiring frequent doses of nebulized medication.[20, 21, 22] Nebulized epinephrine is a component of the Bristol Paediatric Early Warning Tool.[23] The number of medications administered in the preceding 24 hours was included in an early version of the Bedside Paediatric Early Warning System Score.[24] Adding IV antibiotics to the Maximum Modified Early Warning Score improved prediction of the need for higher care utilization among hospitalized adults.[25]

In order to determine the role of the IV medications we found to be associated with clinical deterioration, the necessary next step is to develop a multivariable predictive model to determine if they improve the performance of existing early warning scores in identifying deteriorating patients. Although simplicity is an important characteristic of hand‐calculated early warning scores, integration of a more complex scoring system with more variables, such as these medications, into the electronic health record would allow for automated scoring, eliminating the need to sacrifice score performance to keep the tool simple. Integration into the electronic health record would have the additional benefit of making the score available to clinicians who are not at the bedside. Such tools would be especially useful for remote surveillance for deterioration by critical‐care outreach or rover teams.

Our study has several limitations. First, the sample size was small, and although we sought to minimize the likelihood of chance associations by performing sensitivity analysis, these findings should be confirmed in a larger study. Second, we only evaluated IV medications. Medications administered by other routes could also be associated with clinical deterioration and should be analyzed in future studies. Third, we excluded children hospitalized for <24 hours, as well as transfers that did not meet urgent criteria. These may be limitations because (1) the first 24 hours of hospitalization may be a high‐risk period, and (2) patients who were on trajectories toward severe deterioration and received interventions that prevented further deterioration but did not meet urgent transfer criteria were excluded. It may be that the children we included as cases were at increased risk of deterioration that is either more difficult to recognize early, or more difficult to treat effectively without ICU interventions. Finally, we acknowledge that in some cases the therapeutic classes were associated with deterioration in a causal fashion, and in others the medications administered did not cause deterioration but were signs of therapeutic interventions that were initiated in response to clinical worsening. Identifying the specific indications for administration of drugs used in response to clinical worsening may have resulted in stronger associations with deterioration. However, these indications are often complex, multifactorial, and poorly documented in real time. This limits the ability to automate their detection using the electronic health record, the ultimate goal of this line of research.

CONCLUSION

We used a case‐crossover approach to identify therapeutic classes that are associated with increased risk of clinical deterioration in hospitalized children on pediatric wards. These sentinel therapeutic classes may serve as useful components of electronic health recordbased surveillance tools to detect signs of early, evolving deterioration and flag at‐risk patients for critical‐care outreach or rover team review. Future research should focus on evaluating whether including these therapeutic classes in early warning scores improves their accuracy in detecting signs of deterioration and determining if providing this information as clinical decision support improves patient outcomes.

In recent years, many hospitals have implemented rapid response systems (RRSs) in efforts to reduce mortality outside the intensive care unit (ICU). Rapid response systems include 2 clinical components (efferent and afferent limbs) and 2 organizational components (process improvement and administrative limbs).[1, 2] The efferent limb includes medical emergency teams (METs) that can be summoned to hospital wards to rescue deteriorating patients. The afferent limb identifies patients at risk of deterioration using tools such as early warning scores and triggers a MET response when appropriate.[2] The process‐improvement limb evaluates and optimizes the RRS. The administrative limb implements the RRS and supports its ongoing operation. The effectiveness of most RRSs depends upon the ward team making the decision to escalate care by activating the MET. Barriers to activating the MET may include reduced situational awareness,[3, 4] hierarchical barriers to calling for help,[3, 4, 5, 6, 7, 8] fear of criticism,[3, 8, 9] and other hospital safety cultural barriers.[3, 4, 8]

Proactive critical‐care outreach[10, 11, 12, 13] or rover[14] teams seek to reduce barriers to activation and improve outcomes by systematically identifying and evaluating at‐risk patients without relying on requests for assistance from the ward team. Structured similarly to early warning scores, surveillance tools intended for rover teams might improve their ability to rapidly identify at‐risk patients throughout a hospital. They could combine vital signs with other variables, such as diagnostic and therapeutic interventions that reflect the ward team's early, evolving concern. In particular, the incorporation of medications associated with deterioration may enhance the performance of surveillance tools.

Medications may be associated with deterioration in one of several ways. They could play a causal role in deterioration (ie, opioids causing respiratory insufficiency), represent clinical worsening and anticipation of possible deterioration (ie, broad‐spectrum antibiotics for a positive blood culture), or represent rescue therapies for early deterioration (ie, antihistamines for allergic reactions). In each case, the associated therapeutic classes could be considered sentinel markers of clinical deterioration.

Combined with vital signs and other risk factors, therapeutic classes could serve as useful components of surveillance tools to detect signs of early, evolving deterioration and flag at‐risk patients for evaluation. As a first step, we sought to identify therapeutic classes associated with clinical deterioration. This effort to improve existing afferent tools falls within the process‐improvement limb of RRSs.

PATIENTS AND METHODS

Study Design

We performed a case‐crossover study of children who experienced clinical deterioration. An alternative to the matched case‐control design, the case‐crossover design involves longitudinal within‐subject comparisons exclusively of case subjects such that an individual serves as his or her own control. It is most effective when studying intermittent exposures that result in transient changes in the risk of an acute event,[15, 16, 17] making it appropriate for our study.

Using the case‐crossover design, we compared a discrete time period in close proximity to the deterioration event, called the hazard interval, with earlier time periods in the hospitalization, called the control intervals.[15, 16, 17] In our primary analysis (Figure 1B), we defined the durations of these intervals as follows: We first censored the 2 hours immediately preceding the clinical deterioration event (hours 0 to 2). We made this decision a priori to exclude medications used after deterioration was recognized and resuscitation had already begun. The 12‐hour period immediately preceding the censored interval was the hazard interval (hours 2 to 14). Each 12‐hour period immediately preceding the hazard interval was a control interval (hours 14 to 26, 26 to 38, 38 to 50, and 50 to 62). Depending on the child's length of stay prior to the deterioration event, each hazard interval had 14 control intervals for comparison. In sensitivity analysis, we altered the durations of these intervals (see below).

Figure 1

RESULTS

We identified 12 CPAs, 41 ARCs, and 699 ICU transfers during the study period. Of these 752 events, 141 (19%) were eligible as cases according to our inclusion criteria.[18] (A flowchart demonstrating the identification of eligible cases is provided in Supporting Table 25 in the online version of this article.) Of the 81% excluded, 37% were ICU transfers who did not meet urgent criteria. Another 31% were excluded because they were hospitalized for <24 hours at the time of the event, making their analysis in a case‐crossover design using 12‐hour periods impossible. Event characteristics, demographics, and hospitalization characteristics are shown in Table 2.

Sensitivity Analysis

Of the 14 classes that were significant in primary analysis, we carried 9 forward to sensitivity analysis. The 5 that were not carried forward overlapped substantially with other classes that were carried forward. The decision of which overlapping class to carry forward was based upon (1) parsimony and (2) clinical relevance. This is described briefly in the footnotes to Table 3 (see Supporting information in the online version of this article for a full description of this process). Figure 2 presents the odds ratios and their 95% confidence intervals for the sensitivity analysis of each therapeutic class that was significant in primary analysis. Loop diuretics remained significantly associated with deterioration in all 7 iterations. Glycopeptide antibiotics (vancomycin), third‐generation and fourth‐generation cephalosporins, systemic corticosteroids, and benzodiazepines were significant in 6. Anaerobic antibiotics and narcotic analgesics (full opioid agonists) were significant in 5, and aminoglycoside antibiotics and antidotes to hypersensitivity reactions (diphenhydramine) in 4.

Figure 2

DISCUSSION

We identified 9 therapeutic classes which were associated with a 2.5‐fold to 5.8‐fold increased risk of clinical deterioration. The results were robust to sensitivity analysis. Given their temporal association to the deterioration events, these therapeutic classes may serve as sentinels of early deterioration and are candidate variables to combine with vital signs and other risk factors in a surveillance tool for rover teams or an early warning score.

Although most early warning scores intended for use at the bedside are based upon vital signs and clinical observations, a few also include medications. Monaghan's Pediatric Early Warning Score, the basis for many modified scores used in children's hospitals throughout the world, assigns points for children requiring frequent doses of nebulized medication.[20, 21, 22] Nebulized epinephrine is a component of the Bristol Paediatric Early Warning Tool.[23] The number of medications administered in the preceding 24 hours was included in an early version of the Bedside Paediatric Early Warning System Score.[24] Adding IV antibiotics to the Maximum Modified Early Warning Score improved prediction of the need for higher care utilization among hospitalized adults.[25]

In order to determine the role of the IV medications we found to be associated with clinical deterioration, the necessary next step is to develop a multivariable predictive model to determine if they improve the performance of existing early warning scores in identifying deteriorating patients. Although simplicity is an important characteristic of hand‐calculated early warning scores, integration of a more complex scoring system with more variables, such as these medications, into the electronic health record would allow for automated scoring, eliminating the need to sacrifice score performance to keep the tool simple. Integration into the electronic health record would have the additional benefit of making the score available to clinicians who are not at the bedside. Such tools would be especially useful for remote surveillance for deterioration by critical‐care outreach or rover teams.

Our study has several limitations. First, the sample size was small, and although we sought to minimize the likelihood of chance associations by performing sensitivity analysis, these findings should be confirmed in a larger study. Second, we only evaluated IV medications. Medications administered by other routes could also be associated with clinical deterioration and should be analyzed in future studies. Third, we excluded children hospitalized for <24 hours, as well as transfers that did not meet urgent criteria. These may be limitations because (1) the first 24 hours of hospitalization may be a high‐risk period, and (2) patients who were on trajectories toward severe deterioration and received interventions that prevented further deterioration but did not meet urgent transfer criteria were excluded. It may be that the children we included as cases were at increased risk of deterioration that is either more difficult to recognize early, or more difficult to treat effectively without ICU interventions. Finally, we acknowledge that in some cases the therapeutic classes were associated with deterioration in a causal fashion, and in others the medications administered did not cause deterioration but were signs of therapeutic interventions that were initiated in response to clinical worsening. Identifying the specific indications for administration of drugs used in response to clinical worsening may have resulted in stronger associations with deterioration. However, these indications are often complex, multifactorial, and poorly documented in real time. This limits the ability to automate their detection using the electronic health record, the ultimate goal of this line of research.

CONCLUSION

We used a case‐crossover approach to identify therapeutic classes that are associated with increased risk of clinical deterioration in hospitalized children on pediatric wards. These sentinel therapeutic classes may serve as useful components of electronic health recordbased surveillance tools to detect signs of early, evolving deterioration and flag at‐risk patients for critical‐care outreach or rover team review. Future research should focus on evaluating whether including these therapeutic classes in early warning scores improves their accuracy in detecting signs of deterioration and determining if providing this information as clinical decision support improves patient outcomes.

In‐hospital cardiac arrest (IHCA) research often relies on the first documented cardiac rhythm (FDR) on resuscitation records at the time of cardiopulmonary resuscitation (CPR) initiation as a surrogate for arrest etiology.[1] Over 1000 hospitals report the FDR and associated cardiac arrest data to national registries annually.[2, 3] These data are subsequently used to report national IHCA epidemiology, as well as to develop and refine guidelines for in‐hospital resuscitation.[4]

Suspecting that the FDR might represent the later stage of a progressive cardiopulmonary process rather than a sudden dysrhythmia, we sought to compare the first rhythm documented on resuscitation records at the time of CPR initiation with the telemetry rhythm at the time of the code blue call. We hypothesized that the agreement between FDR and telemetry rhythm would be <80% beyond that predicted by chance (kappa<0.8).[5]

METHODS

RESULTS

During the study period, there were 135 code blue calls for urgent assistance among telemetry‐monitored non‐ICU patients. Of the 135 calls, we excluded 4 events (3%) that did not meet the definition of IHCA, 9 events (7%) with missing or uninterpretable data, and 53 events (39%) with unobtainable data due to automatic purging from the telemetry server. Therefore, 69 events in 69 different patients remained for analysis. Twelve of the 69 included arrests that occurred at Wilmington Hospital and 57 at Christiana Hospital. The characteristics of the patients are shown in Table 2.

Of the 69 arrests, we used the telemetry rhythm at minute 0 in 42 patients (61%), minute 1 in 22 patients (32%), and minute 2 in 5 patients (7%). Agreement between telemetry and FDR was 65% (kappa=0.37, 95% confidence interval: 0.17‐0.56) (Table 3). Agreement did not vary significantly by sex, race, hospital, weekday, time of day, or minute used in the analysis. Agreement was not associated with survival to hospital discharge.

Of the 69 IHCA events, the FDRs vs telemetry rhythms at the time of IHCA were: asystole 17% vs 7%, VTA 25% vs 31%, and other organized rhythms 58% vs 62%. Among the 12 events with FDR recorded as asystole, telemetry at the time of the code call was asystole in 3 (25%), VTA in 1 (8%), and other organized rhythms in 8 (67%). Among the 17 events with FDR recorded as VTA, telemetry at the time of the code call was VTA in 12 (71%) and other organized rhythms in 5 (29%). Among the 40 events with FDR recorded as other organized rhythms, telemetry at the time of the code call was asystole in 2 (5%), VTA in 8 (20%), and other organized rhythms in 30 (75%). Among the 8 patients with VTA on telemetry and other organized rhythms on the resuscitation record, the other organized rhythms were documented as PEA (n=6), sinus (n=1), and bradycardia (n=1). Of the 12 patients with VTA on telemetry and on the resuscitation record, 8 (67%) had ventricular tachycardia on telemetry. Four of the 8 (50%) who had ventricular tachycardia on telemetry had deteriorated into ventricular fibrillation by the time the FDR was recorded. Of the 4 who had ventricular fibrillation on telemetry, all had ventricular fibrillation as the FDR on the resuscitation record.

DISCUSSION

These results establish that FDRs often differ from the telemetry rhythms at the time of the code blue call. This is important because national registries such as the American Heart Association's Get with the GuidelinesResuscitation[2] database use the FDR as a surrogate for arrest etiology, and use their findings to report national IHCA outcomes as well as to develop and refine evidence‐based guidelines for in‐hospital resuscitation. Our findings suggest that using the FDR may be an oversimplification of the complex progression of cardiac rhythms that occurs in the periarrest period. Adding preceding telemetry rhythms to the data elements collected may shed additional light on etiology. Furthermore, our results demonstrate that, among adults with VTA or asystole documented upon arrival of the code blue team, other organized rhythms are often present at the time the staff recognized a life‐threatening condition and called for immediate assistance. This suggests that the VTA and asystole FDRs may represent the later stages of progressive cardiopulmonary processes. This is in contrast to out‐of‐hospital cardiac arrests typically attributed to sudden catastrophic dysrhythmias that often progress to asystole unless rapidly defibrillated.[6, 7, 8] Out‐of‐hospital and in‐hospital arrests are likely different (but overlapping) entities that might benefit from different resuscitation strategies.[9, 10] We hypothesize that, for a subset of these patients, progressive respiratory insufficiency and circulatory shockconditions classically associated more strongly with pediatric than adult IHCAmay have been directly responsible for the event.[1] If future research supports the concept that progressive respiratory insufficiency and circulatory shock are responsible for more adult IHCA than previously recognized, more robust monitoring may be indicated for a larger subset of adult patients hospitalized on general wards. This could include pulse oximetry (wave form can be a surrogate for perfusion), respiratory rate, and/or end‐tidal CO₂ monitoring. In addition, if future research confirms that there is a greater distinction between in‐hospital and out‐of‐hospital cardiac arrest etiology, the expert panels that develop resuscitation guidelines should consider including setting of resuscitation as a branch point in future algorithms.

Our study had several limitations. First, the sample size was small due to uninterpretable rhythm strips, and for 39% of the total code events, the telemetry data had already been purged from the system by the time research staff attempted to retrieve it. Although we do not believe that there was any systematic bias to the data analyzed, the possibility cannot be completely excluded. Second, we were constrained by the inability to retrospectively ascertain the presence of pulses to determine if an organized rhythm identified on telemetry tracings was PEA. Thus, we categorized rhythms into large groups. Although this limited the granularity of the rhythm groups, it was a conservative approach that likely biased toward agreement (the opposite direction of our hypothesis). Third, the lack of perfect time synchronization between the telemetry system, wall clocks in the hospital, and wrist watches that may be referenced when documenting resuscitative efforts on the resuscitation record means that the rhythms we used may have reflected physiology after interventions had already commenced. Thus, in some situations, minute 1, 2, or earlier minutes may more accurately reflect the preintervention rhythm. Highly accurate time synchronization should be a central component of future prospective work in this area.

CONCLUSIONS

The FDR had only fair agreement with the telemetry rhythm at the time of the code blue call. Among those with VTA or asystole documented on CPR initiation, telemetry often revealed other organized rhythms present at the time hospital staff recognized a life‐threatening condition. In contrast to out‐of‐hospital cardiac arrest, FDR of asystole was only rarely preceded by VTA, and FDR of VTA was often preceded by an organized rhythm.[8, 11] Future studies should examine antecedent rhythms in combination with respiratory and perfusion status to more precisely determine arrest etiology.

Disclosures

Dr. Zubrow had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. The authors report no conflicts of interest.

In‐hospital cardiac arrest (IHCA) research often relies on the first documented cardiac rhythm (FDR) on resuscitation records at the time of cardiopulmonary resuscitation (CPR) initiation as a surrogate for arrest etiology.[1] Over 1000 hospitals report the FDR and associated cardiac arrest data to national registries annually.[2, 3] These data are subsequently used to report national IHCA epidemiology, as well as to develop and refine guidelines for in‐hospital resuscitation.[4]

Suspecting that the FDR might represent the later stage of a progressive cardiopulmonary process rather than a sudden dysrhythmia, we sought to compare the first rhythm documented on resuscitation records at the time of CPR initiation with the telemetry rhythm at the time of the code blue call. We hypothesized that the agreement between FDR and telemetry rhythm would be <80% beyond that predicted by chance (kappa<0.8).[5]

METHODS

RESULTS

During the study period, there were 135 code blue calls for urgent assistance among telemetry‐monitored non‐ICU patients. Of the 135 calls, we excluded 4 events (3%) that did not meet the definition of IHCA, 9 events (7%) with missing or uninterpretable data, and 53 events (39%) with unobtainable data due to automatic purging from the telemetry server. Therefore, 69 events in 69 different patients remained for analysis. Twelve of the 69 included arrests that occurred at Wilmington Hospital and 57 at Christiana Hospital. The characteristics of the patients are shown in Table 2.

Of the 69 arrests, we used the telemetry rhythm at minute 0 in 42 patients (61%), minute 1 in 22 patients (32%), and minute 2 in 5 patients (7%). Agreement between telemetry and FDR was 65% (kappa=0.37, 95% confidence interval: 0.17‐0.56) (Table 3). Agreement did not vary significantly by sex, race, hospital, weekday, time of day, or minute used in the analysis. Agreement was not associated with survival to hospital discharge.

Of the 69 IHCA events, the FDRs vs telemetry rhythms at the time of IHCA were: asystole 17% vs 7%, VTA 25% vs 31%, and other organized rhythms 58% vs 62%. Among the 12 events with FDR recorded as asystole, telemetry at the time of the code call was asystole in 3 (25%), VTA in 1 (8%), and other organized rhythms in 8 (67%). Among the 17 events with FDR recorded as VTA, telemetry at the time of the code call was VTA in 12 (71%) and other organized rhythms in 5 (29%). Among the 40 events with FDR recorded as other organized rhythms, telemetry at the time of the code call was asystole in 2 (5%), VTA in 8 (20%), and other organized rhythms in 30 (75%). Among the 8 patients with VTA on telemetry and other organized rhythms on the resuscitation record, the other organized rhythms were documented as PEA (n=6), sinus (n=1), and bradycardia (n=1). Of the 12 patients with VTA on telemetry and on the resuscitation record, 8 (67%) had ventricular tachycardia on telemetry. Four of the 8 (50%) who had ventricular tachycardia on telemetry had deteriorated into ventricular fibrillation by the time the FDR was recorded. Of the 4 who had ventricular fibrillation on telemetry, all had ventricular fibrillation as the FDR on the resuscitation record.

DISCUSSION

These results establish that FDRs often differ from the telemetry rhythms at the time of the code blue call. This is important because national registries such as the American Heart Association's Get with the GuidelinesResuscitation[2] database use the FDR as a surrogate for arrest etiology, and use their findings to report national IHCA outcomes as well as to develop and refine evidence‐based guidelines for in‐hospital resuscitation. Our findings suggest that using the FDR may be an oversimplification of the complex progression of cardiac rhythms that occurs in the periarrest period. Adding preceding telemetry rhythms to the data elements collected may shed additional light on etiology. Furthermore, our results demonstrate that, among adults with VTA or asystole documented upon arrival of the code blue team, other organized rhythms are often present at the time the staff recognized a life‐threatening condition and called for immediate assistance. This suggests that the VTA and asystole FDRs may represent the later stages of progressive cardiopulmonary processes. This is in contrast to out‐of‐hospital cardiac arrests typically attributed to sudden catastrophic dysrhythmias that often progress to asystole unless rapidly defibrillated.[6, 7, 8] Out‐of‐hospital and in‐hospital arrests are likely different (but overlapping) entities that might benefit from different resuscitation strategies.[9, 10] We hypothesize that, for a subset of these patients, progressive respiratory insufficiency and circulatory shockconditions classically associated more strongly with pediatric than adult IHCAmay have been directly responsible for the event.[1] If future research supports the concept that progressive respiratory insufficiency and circulatory shock are responsible for more adult IHCA than previously recognized, more robust monitoring may be indicated for a larger subset of adult patients hospitalized on general wards. This could include pulse oximetry (wave form can be a surrogate for perfusion), respiratory rate, and/or end‐tidal CO₂ monitoring. In addition, if future research confirms that there is a greater distinction between in‐hospital and out‐of‐hospital cardiac arrest etiology, the expert panels that develop resuscitation guidelines should consider including setting of resuscitation as a branch point in future algorithms.

Our study had several limitations. First, the sample size was small due to uninterpretable rhythm strips, and for 39% of the total code events, the telemetry data had already been purged from the system by the time research staff attempted to retrieve it. Although we do not believe that there was any systematic bias to the data analyzed, the possibility cannot be completely excluded. Second, we were constrained by the inability to retrospectively ascertain the presence of pulses to determine if an organized rhythm identified on telemetry tracings was PEA. Thus, we categorized rhythms into large groups. Although this limited the granularity of the rhythm groups, it was a conservative approach that likely biased toward agreement (the opposite direction of our hypothesis). Third, the lack of perfect time synchronization between the telemetry system, wall clocks in the hospital, and wrist watches that may be referenced when documenting resuscitative efforts on the resuscitation record means that the rhythms we used may have reflected physiology after interventions had already commenced. Thus, in some situations, minute 1, 2, or earlier minutes may more accurately reflect the preintervention rhythm. Highly accurate time synchronization should be a central component of future prospective work in this area.

CONCLUSIONS

The FDR had only fair agreement with the telemetry rhythm at the time of the code blue call. Among those with VTA or asystole documented on CPR initiation, telemetry often revealed other organized rhythms present at the time hospital staff recognized a life‐threatening condition. In contrast to out‐of‐hospital cardiac arrest, FDR of asystole was only rarely preceded by VTA, and FDR of VTA was often preceded by an organized rhythm.[8, 11] Future studies should examine antecedent rhythms in combination with respiratory and perfusion status to more precisely determine arrest etiology.

Disclosures

Dr. Zubrow had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. The authors report no conflicts of interest.

In‐hospital cardiac arrest (IHCA) research often relies on the first documented cardiac rhythm (FDR) on resuscitation records at the time of cardiopulmonary resuscitation (CPR) initiation as a surrogate for arrest etiology.[1] Over 1000 hospitals report the FDR and associated cardiac arrest data to national registries annually.[2, 3] These data are subsequently used to report national IHCA epidemiology, as well as to develop and refine guidelines for in‐hospital resuscitation.[4]

Suspecting that the FDR might represent the later stage of a progressive cardiopulmonary process rather than a sudden dysrhythmia, we sought to compare the first rhythm documented on resuscitation records at the time of CPR initiation with the telemetry rhythm at the time of the code blue call. We hypothesized that the agreement between FDR and telemetry rhythm would be <80% beyond that predicted by chance (kappa<0.8).[5]

METHODS

RESULTS

During the study period, there were 135 code blue calls for urgent assistance among telemetry‐monitored non‐ICU patients. Of the 135 calls, we excluded 4 events (3%) that did not meet the definition of IHCA, 9 events (7%) with missing or uninterpretable data, and 53 events (39%) with unobtainable data due to automatic purging from the telemetry server. Therefore, 69 events in 69 different patients remained for analysis. Twelve of the 69 included arrests that occurred at Wilmington Hospital and 57 at Christiana Hospital. The characteristics of the patients are shown in Table 2.

Of the 69 arrests, we used the telemetry rhythm at minute 0 in 42 patients (61%), minute 1 in 22 patients (32%), and minute 2 in 5 patients (7%). Agreement between telemetry and FDR was 65% (kappa=0.37, 95% confidence interval: 0.17‐0.56) (Table 3). Agreement did not vary significantly by sex, race, hospital, weekday, time of day, or minute used in the analysis. Agreement was not associated with survival to hospital discharge.

Of the 69 IHCA events, the FDRs vs telemetry rhythms at the time of IHCA were: asystole 17% vs 7%, VTA 25% vs 31%, and other organized rhythms 58% vs 62%. Among the 12 events with FDR recorded as asystole, telemetry at the time of the code call was asystole in 3 (25%), VTA in 1 (8%), and other organized rhythms in 8 (67%). Among the 17 events with FDR recorded as VTA, telemetry at the time of the code call was VTA in 12 (71%) and other organized rhythms in 5 (29%). Among the 40 events with FDR recorded as other organized rhythms, telemetry at the time of the code call was asystole in 2 (5%), VTA in 8 (20%), and other organized rhythms in 30 (75%). Among the 8 patients with VTA on telemetry and other organized rhythms on the resuscitation record, the other organized rhythms were documented as PEA (n=6), sinus (n=1), and bradycardia (n=1). Of the 12 patients with VTA on telemetry and on the resuscitation record, 8 (67%) had ventricular tachycardia on telemetry. Four of the 8 (50%) who had ventricular tachycardia on telemetry had deteriorated into ventricular fibrillation by the time the FDR was recorded. Of the 4 who had ventricular fibrillation on telemetry, all had ventricular fibrillation as the FDR on the resuscitation record.

DISCUSSION

These results establish that FDRs often differ from the telemetry rhythms at the time of the code blue call. This is important because national registries such as the American Heart Association's Get with the GuidelinesResuscitation[2] database use the FDR as a surrogate for arrest etiology, and use their findings to report national IHCA outcomes as well as to develop and refine evidence‐based guidelines for in‐hospital resuscitation. Our findings suggest that using the FDR may be an oversimplification of the complex progression of cardiac rhythms that occurs in the periarrest period. Adding preceding telemetry rhythms to the data elements collected may shed additional light on etiology. Furthermore, our results demonstrate that, among adults with VTA or asystole documented upon arrival of the code blue team, other organized rhythms are often present at the time the staff recognized a life‐threatening condition and called for immediate assistance. This suggests that the VTA and asystole FDRs may represent the later stages of progressive cardiopulmonary processes. This is in contrast to out‐of‐hospital cardiac arrests typically attributed to sudden catastrophic dysrhythmias that often progress to asystole unless rapidly defibrillated.[6, 7, 8] Out‐of‐hospital and in‐hospital arrests are likely different (but overlapping) entities that might benefit from different resuscitation strategies.[9, 10] We hypothesize that, for a subset of these patients, progressive respiratory insufficiency and circulatory shockconditions classically associated more strongly with pediatric than adult IHCAmay have been directly responsible for the event.[1] If future research supports the concept that progressive respiratory insufficiency and circulatory shock are responsible for more adult IHCA than previously recognized, more robust monitoring may be indicated for a larger subset of adult patients hospitalized on general wards. This could include pulse oximetry (wave form can be a surrogate for perfusion), respiratory rate, and/or end‐tidal CO₂ monitoring. In addition, if future research confirms that there is a greater distinction between in‐hospital and out‐of‐hospital cardiac arrest etiology, the expert panels that develop resuscitation guidelines should consider including setting of resuscitation as a branch point in future algorithms.

Our study had several limitations. First, the sample size was small due to uninterpretable rhythm strips, and for 39% of the total code events, the telemetry data had already been purged from the system by the time research staff attempted to retrieve it. Although we do not believe that there was any systematic bias to the data analyzed, the possibility cannot be completely excluded. Second, we were constrained by the inability to retrospectively ascertain the presence of pulses to determine if an organized rhythm identified on telemetry tracings was PEA. Thus, we categorized rhythms into large groups. Although this limited the granularity of the rhythm groups, it was a conservative approach that likely biased toward agreement (the opposite direction of our hypothesis). Third, the lack of perfect time synchronization between the telemetry system, wall clocks in the hospital, and wrist watches that may be referenced when documenting resuscitative efforts on the resuscitation record means that the rhythms we used may have reflected physiology after interventions had already commenced. Thus, in some situations, minute 1, 2, or earlier minutes may more accurately reflect the preintervention rhythm. Highly accurate time synchronization should be a central component of future prospective work in this area.

CONCLUSIONS

The FDR had only fair agreement with the telemetry rhythm at the time of the code blue call. Among those with VTA or asystole documented on CPR initiation, telemetry often revealed other organized rhythms present at the time hospital staff recognized a life‐threatening condition. In contrast to out‐of‐hospital cardiac arrest, FDR of asystole was only rarely preceded by VTA, and FDR of VTA was often preceded by an organized rhythm.[8, 11] Future studies should examine antecedent rhythms in combination with respiratory and perfusion status to more precisely determine arrest etiology.

Disclosures

Dr. Zubrow had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. The authors report no conflicts of interest.

Thousands of hospitals have implemented rapid response systems in recent years in attempts to reduce mortality outside the intensive care unit (ICU).1 These systems have 2 components, a response arm and an identification arm. The response arm is usually comprised of a multidisciplinary critical care team that responds to calls for urgent assistance outside the ICU; this team is often called a rapid response team or a medical emergency team. The identification arm comes in 2 forms, predictive and detective. Predictive tools estimate a patient's risk of deterioration over time based on factors that are not rapidly changing, such as elements of the patient's history. In contrast, detective tools include highly time‐varying signs of active deterioration, such as vital sign abnormalities.2 To date, most pediatric studies have focused on developing detective tools, including several early warning scores.38

In this study, we sought to identify the characteristics that increase the probability that a hospitalized child will deteriorate, and combine these characteristics into a predictive score. Tools like this may be helpful in identifying and triaging the subset of high‐risk children who should be intensively monitored for early signs of deterioration at the time of admission, as well as in identifying very low‐risk children who, in the absence of other clinical concerns, may be monitored less intensively.

METHODS

Detailed methods, including the inclusion/exclusion criteria, the matching procedures, and a full description of the statistical analysis are provided as an appendix (see Supporting Online Appendix: Supplement to Methods Section in the online version of this article). An abbreviated version follows.

RESULTS

DISCUSSION

Despite the widespread adoption of rapid response systems, we know little about the optimal methods to identify patients whose clinical characteristics alone put them at increased risk of deterioration, and triage the care they receive based on this risk. Pediatric case series have suggested that younger children and those with chronic illnesses are more likely to require assistance from a medical emergency team,1112 but this is the first study to measure their association with this outcome in children.

Most studies with the objective of identifying patients at risk have focused on tools designed to detect symptoms of deterioration that have already begun, using single‐parameter medical emergency team calling criteria1316 or multi‐parameter early warning scores.38 Rather than create a tool to detect deterioration that has already begun, we developed a predictive score that incorporates patient characteristics independently associated with deterioration in hospitalized children, including age <1 year, epilepsy, congenital/genetic defects, history of transplant, enteral tube, hemoglobin <10 g/dL, and blood culture drawn in the preceding 72 hours. The score has the potential to help clinicians identify the children at highest risk of deterioration who might benefit most from the use of vital sign‐based methods to detect deterioration, as well as the children at lowest risk for whom monitoring may be unnecessary. For example, this score could be performed at the time of admission, and those at very low risk of deterioration and without other clinically concerning findings might be considered for a low‐intensity schedule of vital signs and monitoring (such as vital signs every 8 hours, no continuous cardiorespiratory monitoring or pulse oximetry, and early warning score calculation daily), while patients in the intermediate and high‐risk groups might be considered for a more intensive schedule of vital signs and monitoring (such as vital signs every 4 hours, continuous cardiorespiratory monitoring and pulse oximetry, and early warning score calculation every 4 hours). It should be noted, however, that 37 cases (26%) fell into the very low‐risk category, raising the importance of external validation at the point of admission from the emergency department, before the score can be implemented for the potential clinical use described above. If the score performs well in validation studies, then its use in tailoring monitoring parameters has the potential to reduce the amount of time nurses spend responding to false monitor alarms and calculating early warning scores on patients at very low risk of deterioration.

Of note, we excluded children hospitalized for fewer than 24 hours, resulting in the exclusion of 31% of the potentially eligible events. We also excluded 40% of the potentially eligible ICU transfers because they did not meet urgent criteria. These may be limitations because: (1) the first 24 hours of hospitalization may be a high‐risk period; and (2) patients who were on trajectories toward severe deterioration and received interventions that prevented further deterioration, but did not meet urgent criteria, were excluded. It may be that the children we included as cases were at increased risk of deterioration that is either more difficult to recognize early, or more difficult to treat effectively without ICU interventions. In addition, the population of patients meeting urgent criteria may vary across hospitals, limiting generalizability of this score.

In summary, we developed a predictive score and risk stratification tool that may be useful in triaging the intensity of monitoring and surveillance for deterioration that children receive when hospitalized on non‐ICU units. External validation using the setting and frequency of score measurement that would be most valuable clinically (for example, in the emergency department at the time of admission) is needed before clinical use can be recommended.

Thousands of hospitals have implemented rapid response systems in recent years in attempts to reduce mortality outside the intensive care unit (ICU).1 These systems have 2 components, a response arm and an identification arm. The response arm is usually comprised of a multidisciplinary critical care team that responds to calls for urgent assistance outside the ICU; this team is often called a rapid response team or a medical emergency team. The identification arm comes in 2 forms, predictive and detective. Predictive tools estimate a patient's risk of deterioration over time based on factors that are not rapidly changing, such as elements of the patient's history. In contrast, detective tools include highly time‐varying signs of active deterioration, such as vital sign abnormalities.2 To date, most pediatric studies have focused on developing detective tools, including several early warning scores.38

In this study, we sought to identify the characteristics that increase the probability that a hospitalized child will deteriorate, and combine these characteristics into a predictive score. Tools like this may be helpful in identifying and triaging the subset of high‐risk children who should be intensively monitored for early signs of deterioration at the time of admission, as well as in identifying very low‐risk children who, in the absence of other clinical concerns, may be monitored less intensively.

METHODS

Detailed methods, including the inclusion/exclusion criteria, the matching procedures, and a full description of the statistical analysis are provided as an appendix (see Supporting Online Appendix: Supplement to Methods Section in the online version of this article). An abbreviated version follows.

RESULTS