The Centers for Disease Control and Prevention (CDC) established a case definition during the 2009 influenza season for influenza‐like illness (ILI): a temperature of 100.0F (37.8C), oral or equivalent, and cough or sore throat, in the absence of a known cause other than influenza.1 Even though this definition is currently being revised, fever is still listed as a predominant symptom for influenza on the CDC Web site.2, 3

During the spring of 2009 in New York City, 95% of patients with pandemic hemagglutinin type 1 and neuraminidase type 1 (H1N1) influenza A met the case definition for ILI,4 whereas a report from Mexico5 and another from China6 reported that about one‐third of patients presented without fever. A recent study found that the clinical features of H1N1/09 influenza and seasonal strains were similar in hospitalized patients.7 Until recently, the CDC case definition for ILI included fever as a requisite symptom. Most clinicians still consider influenza a fever‐based illness.8, 9 Some studies have reported that seeking early medical care and early treatment for influenza resulted in reduced severity and duration of illness.1013 However, individuals without fever may not have sought early care because their illnesses did not meet all the criteria for ILI.

We hypothesized that significant number of our patients hospitalized during the H1N1/09 influenza pandemic did not have fever as part of the influenza‐like illness prior to hospital presentation. We further hypothesized that this absence of fever was associated with delayed presentation to the hospital and delayed treatment. No prior studies have assessed the effect of fever on hospital presentation, diagnosis, and treatment of H1N1/09 patients. We performed a retrospective cohort analysis comparing H1N1/09 patients with fever as part of their illness prior to hospital presentation to those without fever, to assess the influence of fever on hospital presentation and on diagnosis and treatment in hospitalized patients with laboratory‐confirmed H1N1/09 infection.

METHODS

RESULTS

We identified 135 hospitalized patients with laboratory‐confirmed pandemic H1N1 virus infection during the study period. A substantial number of patients (n = 56, 42%) had no fever as part of their illness prior to hospital presentation. Thirty‐one patients (23%) required treatment in an ICU during hospitalization, and 9 (7%) died. While we observed no differences between patients with fever and those without fever prior to hospital presentation in gender, race, or age (Table 1), patients without fever had higher Charlson index compared to those with fever (P = 0.01, Wilcoxon rank‐sum test). Patients without fever also had significantly longer time to presentation to the hospital since the onset of illness, compared to those with fever (median of 4 vs 2 days, P < 0.001). Patients without fever also had significantly longer time to initiation of proper treatment since the onset of illness, compared to those with fever (median of 5 vs 2 days, P = 0.001).

Patients were rapidly diagnosed and treated after presentation to the hospital with no significant difference by fever status at presentation (P = 0.17 and 0.07 for differences in time from hospital presentation to diagnosis and treatment, respectively). While 82% (65/79) of patients with fever received empiric antibiotic treatment, only 59% (33/56) of those with no fever received empiric antibiotic treatment (P = 0.003). Prevalence of bacterial coinfection was similar between the 2 groups (5% in the fever group vs 9% in the group without fever; P = 0.067). Patients without fever were more frequently in an ICU than those with fever (P = 0.01), but we observed no significant differences in mortality by fever status at presentation in the small number of deaths during the study period (P = 0.61). Compared to patients with no fever, after adjustment for age (<40 vs 40) and Charlson index (0, 12, 3), patients with fever had significantly decreased likelihood of late hospital presentation (>2 days from the onset of illness) (adjusted odds ratio [OR]: 0.27, P = 0.001). Patients with fever also had decreased likelihood of ICU stay compared to patients with no fever, after adjustment for age and Charlson index (adjusted OR: 0.42, P = 0.05) (Table 2).

Patients who presented to the hospital more than 2 days after the onset of illness were significantly more likely to be treated in an ICU during their hospitalization (P < 0.001). Among 71 patients who presented to the hospital more than 2 days after the onset of illness, 27 patients were treated in an ICU and 44 patients did not require ICU care. Among 64 patients who presented to the hospital within 2 days of the onset of illness, 4 patients were treated in an ICU and 60 patients did not require ICU care. Those who presented to the hospital more than 2 days after the onset of illness also had a nonsignificantly higher death rate: 7 of 71 (10%) versus 2 of 64 (3%). Among the comorbidities tested, patients with no fever were significantly more likely to have diabetes mellitus (P = 0.03).

ICU patients had a slightly higher median Charlson index than those who did not receive ICU care (median score 1 vs score 0 for ICU patients vs non‐ICU patients, respectively; P = 0.052). We observed no difference in median Charlson index between patients with late hospital presentation (>2 days from onset of illness) and those with early hospital presentation (2 days from onset of illness) (P = 0.51).

DISCUSSION

We present the first study, to our knowledge, to assess the influence of presence of fever, as part of the illness prior to hospital presentation, on time to presentation and treatment for H1N1/09 patients. By comparing patients who had fever with those who did not, we showed that patients with fever tended to seek medical care sooner and thus received treatment sooner. Those without fever may not have realized they needed to seek medical care until much later in the course of disease compared to those with fever. Given that ICU stay was significantly more likely in patients with late presentation and no fever, these findings have important medical and economic consequences for individual patients and healthcare systems. Our findings, in conjunction with previous studies indicating reduced disease severity with early treatment of influenza,1013, can be used to educate the public to seek medical care for influenza early, even in the absence of fever.

Our study confirms previous observations that a significant number of patients with H1N1/09 did not have fever.5, 6 Nevertheless, fever appeared to trigger initiation of empiric antibiotic treatment. Empiric antibiotic use was significantly higher in the fever group. There was no difference in the median Charlson index among those who received antibiotic treatment and those who did not, suggesting that comorbidities are unlikely to explain the observed association. Empiric antibiotic treatment appeared not to affect whether or not a patient required ICU care during the hospitalization, though this association is difficult to measure because the majority of ICU patients (84%) were admitted to an ICU on the same day as hospital admission. We saw no difference in laboratory‐confirmed bacterial coinfection between patients presenting with and without fever. We did not observe a significant effect of late presentation or fever on mortality, though the small number of deaths in the study likely limited our ability to detect any association.

Recent studies have shown that clinical signs cannot reliably differentiate H1N1/09‐positive and H1N1/09‐negative patients.18, 19 In our study, patients in the fever and non‐fever groups received prompt diagnosis and treatment once they presented to the hospital. This may be because clinicians were encouraged to test patients based on any suspicion for influenza infection and to treat for influenza while waiting for the result of testing. This may explain the similar timeframes for diagnosis and treatment between the fever and non‐fever groups.

The Centers for Disease Control and Prevention (CDC) established a case definition during the 2009 influenza season for influenza‐like illness (ILI): a temperature of 100.0F (37.8C), oral or equivalent, and cough or sore throat, in the absence of a known cause other than influenza.1 Even though this definition is currently being revised, fever is still listed as a predominant symptom for influenza on the CDC Web site.2, 3

During the spring of 2009 in New York City, 95% of patients with pandemic hemagglutinin type 1 and neuraminidase type 1 (H1N1) influenza A met the case definition for ILI,4 whereas a report from Mexico5 and another from China6 reported that about one‐third of patients presented without fever. A recent study found that the clinical features of H1N1/09 influenza and seasonal strains were similar in hospitalized patients.7 Until recently, the CDC case definition for ILI included fever as a requisite symptom. Most clinicians still consider influenza a fever‐based illness.8, 9 Some studies have reported that seeking early medical care and early treatment for influenza resulted in reduced severity and duration of illness.1013 However, individuals without fever may not have sought early care because their illnesses did not meet all the criteria for ILI.

We hypothesized that significant number of our patients hospitalized during the H1N1/09 influenza pandemic did not have fever as part of the influenza‐like illness prior to hospital presentation. We further hypothesized that this absence of fever was associated with delayed presentation to the hospital and delayed treatment. No prior studies have assessed the effect of fever on hospital presentation, diagnosis, and treatment of H1N1/09 patients. We performed a retrospective cohort analysis comparing H1N1/09 patients with fever as part of their illness prior to hospital presentation to those without fever, to assess the influence of fever on hospital presentation and on diagnosis and treatment in hospitalized patients with laboratory‐confirmed H1N1/09 infection.

METHODS

RESULTS

We identified 135 hospitalized patients with laboratory‐confirmed pandemic H1N1 virus infection during the study period. A substantial number of patients (n = 56, 42%) had no fever as part of their illness prior to hospital presentation. Thirty‐one patients (23%) required treatment in an ICU during hospitalization, and 9 (7%) died. While we observed no differences between patients with fever and those without fever prior to hospital presentation in gender, race, or age (Table 1), patients without fever had higher Charlson index compared to those with fever (P = 0.01, Wilcoxon rank‐sum test). Patients without fever also had significantly longer time to presentation to the hospital since the onset of illness, compared to those with fever (median of 4 vs 2 days, P < 0.001). Patients without fever also had significantly longer time to initiation of proper treatment since the onset of illness, compared to those with fever (median of 5 vs 2 days, P = 0.001).

Patients were rapidly diagnosed and treated after presentation to the hospital with no significant difference by fever status at presentation (P = 0.17 and 0.07 for differences in time from hospital presentation to diagnosis and treatment, respectively). While 82% (65/79) of patients with fever received empiric antibiotic treatment, only 59% (33/56) of those with no fever received empiric antibiotic treatment (P = 0.003). Prevalence of bacterial coinfection was similar between the 2 groups (5% in the fever group vs 9% in the group without fever; P = 0.067). Patients without fever were more frequently in an ICU than those with fever (P = 0.01), but we observed no significant differences in mortality by fever status at presentation in the small number of deaths during the study period (P = 0.61). Compared to patients with no fever, after adjustment for age (<40 vs 40) and Charlson index (0, 12, 3), patients with fever had significantly decreased likelihood of late hospital presentation (>2 days from the onset of illness) (adjusted odds ratio [OR]: 0.27, P = 0.001). Patients with fever also had decreased likelihood of ICU stay compared to patients with no fever, after adjustment for age and Charlson index (adjusted OR: 0.42, P = 0.05) (Table 2).

Patients who presented to the hospital more than 2 days after the onset of illness were significantly more likely to be treated in an ICU during their hospitalization (P < 0.001). Among 71 patients who presented to the hospital more than 2 days after the onset of illness, 27 patients were treated in an ICU and 44 patients did not require ICU care. Among 64 patients who presented to the hospital within 2 days of the onset of illness, 4 patients were treated in an ICU and 60 patients did not require ICU care. Those who presented to the hospital more than 2 days after the onset of illness also had a nonsignificantly higher death rate: 7 of 71 (10%) versus 2 of 64 (3%). Among the comorbidities tested, patients with no fever were significantly more likely to have diabetes mellitus (P = 0.03).

ICU patients had a slightly higher median Charlson index than those who did not receive ICU care (median score 1 vs score 0 for ICU patients vs non‐ICU patients, respectively; P = 0.052). We observed no difference in median Charlson index between patients with late hospital presentation (>2 days from onset of illness) and those with early hospital presentation (2 days from onset of illness) (P = 0.51).

DISCUSSION

We present the first study, to our knowledge, to assess the influence of presence of fever, as part of the illness prior to hospital presentation, on time to presentation and treatment for H1N1/09 patients. By comparing patients who had fever with those who did not, we showed that patients with fever tended to seek medical care sooner and thus received treatment sooner. Those without fever may not have realized they needed to seek medical care until much later in the course of disease compared to those with fever. Given that ICU stay was significantly more likely in patients with late presentation and no fever, these findings have important medical and economic consequences for individual patients and healthcare systems. Our findings, in conjunction with previous studies indicating reduced disease severity with early treatment of influenza,1013, can be used to educate the public to seek medical care for influenza early, even in the absence of fever.

Our study confirms previous observations that a significant number of patients with H1N1/09 did not have fever.5, 6 Nevertheless, fever appeared to trigger initiation of empiric antibiotic treatment. Empiric antibiotic use was significantly higher in the fever group. There was no difference in the median Charlson index among those who received antibiotic treatment and those who did not, suggesting that comorbidities are unlikely to explain the observed association. Empiric antibiotic treatment appeared not to affect whether or not a patient required ICU care during the hospitalization, though this association is difficult to measure because the majority of ICU patients (84%) were admitted to an ICU on the same day as hospital admission. We saw no difference in laboratory‐confirmed bacterial coinfection between patients presenting with and without fever. We did not observe a significant effect of late presentation or fever on mortality, though the small number of deaths in the study likely limited our ability to detect any association.

Recent studies have shown that clinical signs cannot reliably differentiate H1N1/09‐positive and H1N1/09‐negative patients.18, 19 In our study, patients in the fever and non‐fever groups received prompt diagnosis and treatment once they presented to the hospital. This may be because clinicians were encouraged to test patients based on any suspicion for influenza infection and to treat for influenza while waiting for the result of testing. This may explain the similar timeframes for diagnosis and treatment between the fever and non‐fever groups.

The Centers for Disease Control and Prevention (CDC) established a case definition during the 2009 influenza season for influenza‐like illness (ILI): a temperature of 100.0F (37.8C), oral or equivalent, and cough or sore throat, in the absence of a known cause other than influenza.1 Even though this definition is currently being revised, fever is still listed as a predominant symptom for influenza on the CDC Web site.2, 3

During the spring of 2009 in New York City, 95% of patients with pandemic hemagglutinin type 1 and neuraminidase type 1 (H1N1) influenza A met the case definition for ILI,4 whereas a report from Mexico5 and another from China6 reported that about one‐third of patients presented without fever. A recent study found that the clinical features of H1N1/09 influenza and seasonal strains were similar in hospitalized patients.7 Until recently, the CDC case definition for ILI included fever as a requisite symptom. Most clinicians still consider influenza a fever‐based illness.8, 9 Some studies have reported that seeking early medical care and early treatment for influenza resulted in reduced severity and duration of illness.1013 However, individuals without fever may not have sought early care because their illnesses did not meet all the criteria for ILI.

We hypothesized that significant number of our patients hospitalized during the H1N1/09 influenza pandemic did not have fever as part of the influenza‐like illness prior to hospital presentation. We further hypothesized that this absence of fever was associated with delayed presentation to the hospital and delayed treatment. No prior studies have assessed the effect of fever on hospital presentation, diagnosis, and treatment of H1N1/09 patients. We performed a retrospective cohort analysis comparing H1N1/09 patients with fever as part of their illness prior to hospital presentation to those without fever, to assess the influence of fever on hospital presentation and on diagnosis and treatment in hospitalized patients with laboratory‐confirmed H1N1/09 infection.

METHODS

RESULTS

We identified 135 hospitalized patients with laboratory‐confirmed pandemic H1N1 virus infection during the study period. A substantial number of patients (n = 56, 42%) had no fever as part of their illness prior to hospital presentation. Thirty‐one patients (23%) required treatment in an ICU during hospitalization, and 9 (7%) died. While we observed no differences between patients with fever and those without fever prior to hospital presentation in gender, race, or age (Table 1), patients without fever had higher Charlson index compared to those with fever (P = 0.01, Wilcoxon rank‐sum test). Patients without fever also had significantly longer time to presentation to the hospital since the onset of illness, compared to those with fever (median of 4 vs 2 days, P < 0.001). Patients without fever also had significantly longer time to initiation of proper treatment since the onset of illness, compared to those with fever (median of 5 vs 2 days, P = 0.001).

Patients were rapidly diagnosed and treated after presentation to the hospital with no significant difference by fever status at presentation (P = 0.17 and 0.07 for differences in time from hospital presentation to diagnosis and treatment, respectively). While 82% (65/79) of patients with fever received empiric antibiotic treatment, only 59% (33/56) of those with no fever received empiric antibiotic treatment (P = 0.003). Prevalence of bacterial coinfection was similar between the 2 groups (5% in the fever group vs 9% in the group without fever; P = 0.067). Patients without fever were more frequently in an ICU than those with fever (P = 0.01), but we observed no significant differences in mortality by fever status at presentation in the small number of deaths during the study period (P = 0.61). Compared to patients with no fever, after adjustment for age (<40 vs 40) and Charlson index (0, 12, 3), patients with fever had significantly decreased likelihood of late hospital presentation (>2 days from the onset of illness) (adjusted odds ratio [OR]: 0.27, P = 0.001). Patients with fever also had decreased likelihood of ICU stay compared to patients with no fever, after adjustment for age and Charlson index (adjusted OR: 0.42, P = 0.05) (Table 2).

Patients who presented to the hospital more than 2 days after the onset of illness were significantly more likely to be treated in an ICU during their hospitalization (P < 0.001). Among 71 patients who presented to the hospital more than 2 days after the onset of illness, 27 patients were treated in an ICU and 44 patients did not require ICU care. Among 64 patients who presented to the hospital within 2 days of the onset of illness, 4 patients were treated in an ICU and 60 patients did not require ICU care. Those who presented to the hospital more than 2 days after the onset of illness also had a nonsignificantly higher death rate: 7 of 71 (10%) versus 2 of 64 (3%). Among the comorbidities tested, patients with no fever were significantly more likely to have diabetes mellitus (P = 0.03).

ICU patients had a slightly higher median Charlson index than those who did not receive ICU care (median score 1 vs score 0 for ICU patients vs non‐ICU patients, respectively; P = 0.052). We observed no difference in median Charlson index between patients with late hospital presentation (>2 days from onset of illness) and those with early hospital presentation (2 days from onset of illness) (P = 0.51).

DISCUSSION

We present the first study, to our knowledge, to assess the influence of presence of fever, as part of the illness prior to hospital presentation, on time to presentation and treatment for H1N1/09 patients. By comparing patients who had fever with those who did not, we showed that patients with fever tended to seek medical care sooner and thus received treatment sooner. Those without fever may not have realized they needed to seek medical care until much later in the course of disease compared to those with fever. Given that ICU stay was significantly more likely in patients with late presentation and no fever, these findings have important medical and economic consequences for individual patients and healthcare systems. Our findings, in conjunction with previous studies indicating reduced disease severity with early treatment of influenza,1013, can be used to educate the public to seek medical care for influenza early, even in the absence of fever.

Our study confirms previous observations that a significant number of patients with H1N1/09 did not have fever.5, 6 Nevertheless, fever appeared to trigger initiation of empiric antibiotic treatment. Empiric antibiotic use was significantly higher in the fever group. There was no difference in the median Charlson index among those who received antibiotic treatment and those who did not, suggesting that comorbidities are unlikely to explain the observed association. Empiric antibiotic treatment appeared not to affect whether or not a patient required ICU care during the hospitalization, though this association is difficult to measure because the majority of ICU patients (84%) were admitted to an ICU on the same day as hospital admission. We saw no difference in laboratory‐confirmed bacterial coinfection between patients presenting with and without fever. We did not observe a significant effect of late presentation or fever on mortality, though the small number of deaths in the study likely limited our ability to detect any association.

Recent studies have shown that clinical signs cannot reliably differentiate H1N1/09‐positive and H1N1/09‐negative patients.18, 19 In our study, patients in the fever and non‐fever groups received prompt diagnosis and treatment once they presented to the hospital. This may be because clinicians were encouraged to test patients based on any suspicion for influenza infection and to treat for influenza while waiting for the result of testing. This may explain the similar timeframes for diagnosis and treatment between the fever and non‐fever groups.

There are over 350,000 cases of out‐of‐hospital cardiac arrest (OHCA) each year in the United States1, 2 and, with supportive therapy alone, only a fraction of victims survive to hospital discharge. Rapid intervention including cardiopulmonary resuscitation in the moments following arrest is critical to minimizing neurologic injury, morbidity, and mortality. In 2002, two small randomized controlled trials showed a survival benefit of therapeutic hypothermia (TH) when provided within 12 hours after return of circulation following an OHCA and, to date, TH remains one of the few interventions with proven mortality benefit after initial cardiopulmonary resuscitation.3, 4 Since 2003, TH has been incorporated into the American Heart Association practice guidelines57 and use of TH has steadily increased, but widespread clinical uptake remains low.8, 9

The initial studies that evaluated TH were small, with only 189 patients included in the TH arms of the 2 trials combined. To date, only a few studies have replicated this initial observation in real‐world settings, with little analysis of outcomes in US centers in particular.1013 Accordingly, we aimed to examine the real‐world experience with TH in the United States using a large administrative claims database of all California hospital admissions to describe utilization trends, hospital mortality, and volumeoutcome relationships associated with the intervention.

MATERIALS AND METHODS

RESULTS

Descriptive Data

Table 1 reports summary statistics for patients with cardiac arrest complicated by neurologic insult (anoxic brain injury, coma, or persistent vegetative state) between 1999 and 2008. Across all years, 204 patients were identified as undergoing TH. In comparison, 105 patients were identified as undergoing TH in 2008 alone. Patients who underwent TH were less likely to be male (30.7% vs 44.6% male, P < 0.01), were younger (63.9 15.0 years vs 67.3 15.7 years, P = 0.03), and had equivalent numbers of Charlson‐Deyo comorbidities (2.5 2.0 diagnoses vs 2.5 2.0 diagnoses, P = 0.89). Therapeutic hypothermia was more commonly employed at teaching hospitals (51/3709 [1.4%] vs 153/42,942 [0.4%], P < 0.01). There was a trend toward decreased unadjusted mortality among patients who underwent therapeutic hypothermia compared with those who did not (56.9% vs 62.8%, P = 0.08).

Table 1.Characteristics of Patients Suffering Cardiac Arrest Complicated by Neurologic Insult, Stratified by Therapeutic Hypothermia Use, California Hospitals 19992008

	Therapeutic Hypothermia	No Therapeutic Hypothermia	P Value
P values for age, hospital mortality, and number of Charlson‐Deyo comorbidities reflect 2‐sided t test of continuous variables; P values for all other variables reflect 2 test of categorical data.
No. observations	204	46,629
No. cases in teaching hospitals	51	3,658
No. cases in non‐teaching hospitals	153	42,789
Age, y	63.9 15.0	67.3 15.7	0.06
Male	30.7	44.6	<0.01
Hospital mortality, %	56.9	62.8	0.08
Comorbidities
No. Charlson‐Deyo comorbidities	2.5 2.0	2.5 2.0	0.89
Coronary artery disease, %	48.0	38.0	<0.01
Acute myocardial infarction, %	42.6	28.9	<0.01
Congestive heart failure, %	27.9	35.3	0.03
Hypertension, %	36.3	33.2	0.83
Acute renal failure, %	33.3	26.6	0.03
Diabetes mellitus, %	30.9	23.0	<0.01
Chronic obstructive pulmonary disease, %	10.3	19.3	<0.01

Figures 1 and 2 provide additional aggregate statistics on therapeutic hypothermia in California hospitals. Figure 1 plots the number of therapeutic hypothermia cases recorded in the administrative discharge registry between 1999 and 2008. Of the 204 total cases identified during this period, 178 (87.3%) were performed between 2006 and 2008. Figure 2 shows the distribution of TH cases across centers that performed therapeutic hypothermia (n = 47 hospitals, 11.3 % of all hospitals). Ten centers accounted for 124/204 (60.7%) of the total patients treated with TH after cardiac arrest; the top 3 centers accounted for 64 (31.4%) of the treated patients. Twenty‐seven hospitals were identified as performing therapeutic hypothermia on only 1 or 2 patients between 1999 and 2008.

Figure 1

Figure 2

Risk‐Adjusted Mortality

Table 2 presents the odds ratio of factors predicting in‐hospital mortality after cardiac arrest complicated by neurologic insult. Factors include use of TH after cardiac arrest, age, gender, year of admission, number of Charlson‐Deyo comorbidities, hospital teaching status, and volume tercile of hospitals that performed therapeutic hypothermia. Overall, patients who were older, male, and had greater comorbidities were statistically more likely to die after cardiac arrest complicated by neurologic insult. Regardless of whether they underwent TH, patients admitted to hospitals in the highest volume tercile of TH use were more likely to die after cardiac arrest. Adjusting for volume tercile, teaching hospital status was not independently associated with mortality after cardiac arrest. The adjusted odds ratio of mortality among patients undergoing therapeutic hypothermia was 0.80 (95% confidence interval [CI] 0.601.06, P = 0.11). The adjusted probability of inpatient mortality among patients undergoing therapeutic hypothermia was 57.5% (95% CI 50.764.3%) compared to those who did not 62.8% (95% CI 61.763.9%, P = 0.11).

Table 2.Adjusted Odds Ratio of Hospital Mortality After Cardiac Arrest Complicated by Neurologic Insult in a Multivariable Regression Model, California Hospitals 19992008

Variable	Odds Ratio of Hospital Mortality (95% CI)	P Value
Abbreviations: CI, confidence interval; TH, therapeutic hypothermia. *Odds ratios compared to individuals aged 6065. Odds ratios compared to hospitals not performing TH.
No. observations	46,651
Age*
6569	1.19 (1.121.28)	<0.001
7074	1.29 (1.201.39)	<0.001
7579	1.55 (1.441.67)	<0.001
8084	1.79 (1.651.93)	<0.001
85 and over	2.06 (1.892.25)	<0.001
Male	1.15 (1.101.21)	<0.001
Teaching hospital	1.13 (0.951.34)	0.17
No. Charlson‐Deyo comorbidities	1.09 (1.081.10)	<0.001
Year trend	0.98 (0.970.99)	<0.001
Volume tercile among hospitals performing TH
First tercile	0.94 (0.791.12)	0.48
Second tercile	1.03 (0.801.33)	0.82
Third tercile	1.20 (1.051.36)	0.006
Therapeutic hypothermia	0.80 (0.601.06)	0.11

Figure 3 presents adjusted mortality after cardiac arrest in hospitals that did not perform TH, as well as adjusted mortality associated with TH for each volume tercile of hospitals that performed the procedure. Hospital mortality rates among patients not receiving TH after cardiac arrest were slightly higher in hospitals in the high volume tercile of TH (66.3%, 95% CI 63.868.8%) compared to hospitals in low and moderate volume terciles and to hospitals not performing TH (P < 0.001). Hospital mortality rates among low and moderate TH volume centers and in centers not performing TH were similar (62.3%, 61.3%, and 63.4%, respectively). Among both the low volume and moderate volume terciles, however, patients who underwent TH after cardiac arrest were significantly less likely to die in‐hospital compared to those who did not. For patients admitted to hospitals in the low volume tercile, those undergoing therapeutic hypothermia had an adjusted hospital mortality rate of 25.5% (95% CI 3.047.9%) compared to those who did not undergo TH (adjusted mortality 61.3%, 95% CI 57.465.1%), P < 0.001. In the moderate volume tercile, patients receiving therapeutic hypothermia had an adjusted hospital mortality rate of 31.0% (95% CI 9.2%52.8%) compared to 63.4% (95% CI 57.769.1%), P < 0.001, among those not undergoing the procedure. There was no statistically significant difference in adjusted mortality between those who underwent TH and those who did not, in hospitals in the highest volume tercile (P = 0.211). In addition to examining how volume of therapeutic hypothermia performed by hospitals affected the association between TH and hospital mortality, we also examined whether year of admission and teaching hospital independently modified the association. Neither year of admission nor teaching hospital statistically significantly affected the association between therapeutic hypothermia and hospital mortality after cardiac arrest at the P < 0.10 level.

Figure 3

DISCUSSION

In an administrative database of all admissions to California hospitals, we demonstrated that use of TH increased steadily since the publication of the initial clinical trials in 2002. The absolute level of TH utilization in our study undoubtedly represents a significant underestimation of actual TH utilization, however, our study does provide an assessment of the utilization trends over time. The bulk of TH use appears to be performed in a small group of high volume centers, and 89% of California hospitals did not perform TH during the study period (as assessed by procedure billing codes). Additionally, within the limitations of a retrospective, administrative claims‐based study design, TH appears to be associated with a similar in‐hospital mortality rate to that seen in clinical trials.3, 4 In exploratory analyses, there appears to be a particular benefit of TH in low and moderate volume centers, though these findings should be considered hypothesis‐generating.

Despite the body of evidence in favor of TH, utilization in our study and others appears quite low. In a 2005 survey of physicians, 87% of respondents had never used therapeutic hypothermia, citing inadequate data, technical limitations, and lack of incorporation in the Advanced Cardiac Life Support (ACLS) protocol as principal justifications.8 Other surveys have shown similar results and noted that critical care physicians and those working in large medical centers were more likely to adopt the therapy.9 Advocates of the therapy have suggested that an explicit hospital‐based plan developed by key stakeholders can help facilitate implementation.21 Accordingly, there is growing interest in developing centers of expertise in highly intensive therapies such as TH. For instance, the New York City Emergency Medical Service has begun to explore a protocol to divert TH candidates to specialized centers.22, 23 Some favorable results have been reported in individual hospitals and local hospital systems.2428

Our data suggest that TH is associated with an in‐hospital mortality rate that is comparable to that reported in the clinical trials. For example, in a 2009 meta‐analysis of 4 clinical trials and 1 abstract (481 patients in total), TH was associated with a 35% relative mortality benefit as compared to standard post‐resuscitation care.29 It has been estimated that broad TH implementation could save thousands of lives30 and many authors have advocated for its use and outlined explicit protocols for implementation.17 Furthermore, TH appears to be cost‐effective in line with other accepted therapies. Assuming the Hypothermia After Cardiac Arrest (HACA) trial inclusion criteria, even at extreme estimates for costs, the cost‐effectiveness of hypothermia remains less than $100,000 per quality‐adjusted life year.31

There are important limitations of this study. Our use of administrative claims data certainly underestimates the level of TH utilization, since we could only identify cases in which TH was included in the billing codes for the hospitalization. Hospitals may vary in utilization of this particular billing code for TH in ways that bias our estimated associations. The ICD‐9 code 99.81 for therapeutic hypothermia was also not developed for post‐cardiac arrest TH specifically, so use of the code may actually lag clinical utilization. Although the observed trend in TH utilization is likely mainly due to a true increase in utilization, it is possible that some of the observed increase is due to an increase in utilization of TH procedure billing codes. Our TH utilization estimates should be construed as a lower bound of the actual rates. Additionally, although the estimated real‐world mortality benefit of TH may be comparable to that of clinical trials, the equivalence of patients in our sample to those in published randomized trials is uncertain. Similarly, even after adjusting for age, gender, year of admission, comorbidities, hospital teaching status, and TH volume, there are likely many unmeasured variables that influence mortality in both the TH and comparison groups. There are also likely patients included in our comparison group who had both cardiac arrest or ventricular tachycardia and anoxic brain injury, but who were not candidates for TH as the episode of cardiac arrest followed rather than preceded the anoxic brain injury. Since we lack detailed clinical data about the TH cases (initial rhythm, time before return of circulation, preexisting disease states, etc.), we are unable to match controls directly to cases. Additionally, we lack data to assess neurologic recovery or quality of life after arrest.

The observation that a mortality benefit in our study could be detected only in low and moderate volume centers requires further exploration. Indeed, one might expect that high volume centers may have better outcomes with TH as a result of more robust infrastructure, technical experience, and available resources. Our finding that mortality benefits of TH appear concentrated in centers with low to moderate volume of TH utilization suggest at least 1 of 2 possibilities. First, low and moderate volume centers may perform TH in a subset of patients who benefit most from the intervention or, alternatively, in the most viable cardiac arrest cases (those who may fare well with or without the therapy). Consequently, we may observe relatively favorable outcomes in this group due to this selection bias. Second, high volume centersdespite having more expertisemay also attract patients at higher mortality risk due to referral bias. This would lead us to estimate lower mortality benefits associated with TH in these high volume centers. Indeed, greater observed mortality at high volume centers regardless of TH status suggests that overall acuity is higher at high volume centers. While our inferences are greatly affected by issues of case selection and referral bias, it also important to consider the possibility that the estimated mortality benefit of TH in higher volume centers is lower because of the selection of patients who do not meet current guidelines for treatment with TH. Distinguishing whether the selection of patients undergoing TH at high volume centers is appropriate or inappropriate based on current guidelines is an important issue that merits further research with datasets with more refined patient clinical information.

In summary, therapeutic hypothermia utilization is low, but the rate of implementation has increased since the publication of the initial clinical trials in 2002. The bulk of TH utilization appears limited to a subset of high volume centers, and most centers in California appear to have not used the therapy. Real‐world in‐hospital mortality associated with TH is comparable to that reported in randomized clinical trials.

There are over 350,000 cases of out‐of‐hospital cardiac arrest (OHCA) each year in the United States1, 2 and, with supportive therapy alone, only a fraction of victims survive to hospital discharge. Rapid intervention including cardiopulmonary resuscitation in the moments following arrest is critical to minimizing neurologic injury, morbidity, and mortality. In 2002, two small randomized controlled trials showed a survival benefit of therapeutic hypothermia (TH) when provided within 12 hours after return of circulation following an OHCA and, to date, TH remains one of the few interventions with proven mortality benefit after initial cardiopulmonary resuscitation.3, 4 Since 2003, TH has been incorporated into the American Heart Association practice guidelines57 and use of TH has steadily increased, but widespread clinical uptake remains low.8, 9

The initial studies that evaluated TH were small, with only 189 patients included in the TH arms of the 2 trials combined. To date, only a few studies have replicated this initial observation in real‐world settings, with little analysis of outcomes in US centers in particular.1013 Accordingly, we aimed to examine the real‐world experience with TH in the United States using a large administrative claims database of all California hospital admissions to describe utilization trends, hospital mortality, and volumeoutcome relationships associated with the intervention.

MATERIALS AND METHODS

RESULTS

Descriptive Data

Table 1 reports summary statistics for patients with cardiac arrest complicated by neurologic insult (anoxic brain injury, coma, or persistent vegetative state) between 1999 and 2008. Across all years, 204 patients were identified as undergoing TH. In comparison, 105 patients were identified as undergoing TH in 2008 alone. Patients who underwent TH were less likely to be male (30.7% vs 44.6% male, P < 0.01), were younger (63.9 15.0 years vs 67.3 15.7 years, P = 0.03), and had equivalent numbers of Charlson‐Deyo comorbidities (2.5 2.0 diagnoses vs 2.5 2.0 diagnoses, P = 0.89). Therapeutic hypothermia was more commonly employed at teaching hospitals (51/3709 [1.4%] vs 153/42,942 [0.4%], P < 0.01). There was a trend toward decreased unadjusted mortality among patients who underwent therapeutic hypothermia compared with those who did not (56.9% vs 62.8%, P = 0.08).

Table 1.Characteristics of Patients Suffering Cardiac Arrest Complicated by Neurologic Insult, Stratified by Therapeutic Hypothermia Use, California Hospitals 19992008

	Therapeutic Hypothermia	No Therapeutic Hypothermia	P Value
P values for age, hospital mortality, and number of Charlson‐Deyo comorbidities reflect 2‐sided t test of continuous variables; P values for all other variables reflect 2 test of categorical data.
No. observations	204	46,629
No. cases in teaching hospitals	51	3,658
No. cases in non‐teaching hospitals	153	42,789
Age, y	63.9 15.0	67.3 15.7	0.06
Male	30.7	44.6	<0.01
Hospital mortality, %	56.9	62.8	0.08
Comorbidities
No. Charlson‐Deyo comorbidities	2.5 2.0	2.5 2.0	0.89
Coronary artery disease, %	48.0	38.0	<0.01
Acute myocardial infarction, %	42.6	28.9	<0.01
Congestive heart failure, %	27.9	35.3	0.03
Hypertension, %	36.3	33.2	0.83
Acute renal failure, %	33.3	26.6	0.03
Diabetes mellitus, %	30.9	23.0	<0.01
Chronic obstructive pulmonary disease, %	10.3	19.3	<0.01

Figures 1 and 2 provide additional aggregate statistics on therapeutic hypothermia in California hospitals. Figure 1 plots the number of therapeutic hypothermia cases recorded in the administrative discharge registry between 1999 and 2008. Of the 204 total cases identified during this period, 178 (87.3%) were performed between 2006 and 2008. Figure 2 shows the distribution of TH cases across centers that performed therapeutic hypothermia (n = 47 hospitals, 11.3 % of all hospitals). Ten centers accounted for 124/204 (60.7%) of the total patients treated with TH after cardiac arrest; the top 3 centers accounted for 64 (31.4%) of the treated patients. Twenty‐seven hospitals were identified as performing therapeutic hypothermia on only 1 or 2 patients between 1999 and 2008.

Figure 1

Figure 2

Risk‐Adjusted Mortality

Table 2 presents the odds ratio of factors predicting in‐hospital mortality after cardiac arrest complicated by neurologic insult. Factors include use of TH after cardiac arrest, age, gender, year of admission, number of Charlson‐Deyo comorbidities, hospital teaching status, and volume tercile of hospitals that performed therapeutic hypothermia. Overall, patients who were older, male, and had greater comorbidities were statistically more likely to die after cardiac arrest complicated by neurologic insult. Regardless of whether they underwent TH, patients admitted to hospitals in the highest volume tercile of TH use were more likely to die after cardiac arrest. Adjusting for volume tercile, teaching hospital status was not independently associated with mortality after cardiac arrest. The adjusted odds ratio of mortality among patients undergoing therapeutic hypothermia was 0.80 (95% confidence interval [CI] 0.601.06, P = 0.11). The adjusted probability of inpatient mortality among patients undergoing therapeutic hypothermia was 57.5% (95% CI 50.764.3%) compared to those who did not 62.8% (95% CI 61.763.9%, P = 0.11).

Table 2.Adjusted Odds Ratio of Hospital Mortality After Cardiac Arrest Complicated by Neurologic Insult in a Multivariable Regression Model, California Hospitals 19992008

Variable	Odds Ratio of Hospital Mortality (95% CI)	P Value
Abbreviations: CI, confidence interval; TH, therapeutic hypothermia. *Odds ratios compared to individuals aged 6065. Odds ratios compared to hospitals not performing TH.
No. observations	46,651
Age*
6569	1.19 (1.121.28)	<0.001
7074	1.29 (1.201.39)	<0.001
7579	1.55 (1.441.67)	<0.001
8084	1.79 (1.651.93)	<0.001
85 and over	2.06 (1.892.25)	<0.001
Male	1.15 (1.101.21)	<0.001
Teaching hospital	1.13 (0.951.34)	0.17
No. Charlson‐Deyo comorbidities	1.09 (1.081.10)	<0.001
Year trend	0.98 (0.970.99)	<0.001
Volume tercile among hospitals performing TH
First tercile	0.94 (0.791.12)	0.48
Second tercile	1.03 (0.801.33)	0.82
Third tercile	1.20 (1.051.36)	0.006
Therapeutic hypothermia	0.80 (0.601.06)	0.11

Figure 3 presents adjusted mortality after cardiac arrest in hospitals that did not perform TH, as well as adjusted mortality associated with TH for each volume tercile of hospitals that performed the procedure. Hospital mortality rates among patients not receiving TH after cardiac arrest were slightly higher in hospitals in the high volume tercile of TH (66.3%, 95% CI 63.868.8%) compared to hospitals in low and moderate volume terciles and to hospitals not performing TH (P < 0.001). Hospital mortality rates among low and moderate TH volume centers and in centers not performing TH were similar (62.3%, 61.3%, and 63.4%, respectively). Among both the low volume and moderate volume terciles, however, patients who underwent TH after cardiac arrest were significantly less likely to die in‐hospital compared to those who did not. For patients admitted to hospitals in the low volume tercile, those undergoing therapeutic hypothermia had an adjusted hospital mortality rate of 25.5% (95% CI 3.047.9%) compared to those who did not undergo TH (adjusted mortality 61.3%, 95% CI 57.465.1%), P < 0.001. In the moderate volume tercile, patients receiving therapeutic hypothermia had an adjusted hospital mortality rate of 31.0% (95% CI 9.2%52.8%) compared to 63.4% (95% CI 57.769.1%), P < 0.001, among those not undergoing the procedure. There was no statistically significant difference in adjusted mortality between those who underwent TH and those who did not, in hospitals in the highest volume tercile (P = 0.211). In addition to examining how volume of therapeutic hypothermia performed by hospitals affected the association between TH and hospital mortality, we also examined whether year of admission and teaching hospital independently modified the association. Neither year of admission nor teaching hospital statistically significantly affected the association between therapeutic hypothermia and hospital mortality after cardiac arrest at the P < 0.10 level.

Figure 3

DISCUSSION

In an administrative database of all admissions to California hospitals, we demonstrated that use of TH increased steadily since the publication of the initial clinical trials in 2002. The absolute level of TH utilization in our study undoubtedly represents a significant underestimation of actual TH utilization, however, our study does provide an assessment of the utilization trends over time. The bulk of TH use appears to be performed in a small group of high volume centers, and 89% of California hospitals did not perform TH during the study period (as assessed by procedure billing codes). Additionally, within the limitations of a retrospective, administrative claims‐based study design, TH appears to be associated with a similar in‐hospital mortality rate to that seen in clinical trials.3, 4 In exploratory analyses, there appears to be a particular benefit of TH in low and moderate volume centers, though these findings should be considered hypothesis‐generating.

Despite the body of evidence in favor of TH, utilization in our study and others appears quite low. In a 2005 survey of physicians, 87% of respondents had never used therapeutic hypothermia, citing inadequate data, technical limitations, and lack of incorporation in the Advanced Cardiac Life Support (ACLS) protocol as principal justifications.8 Other surveys have shown similar results and noted that critical care physicians and those working in large medical centers were more likely to adopt the therapy.9 Advocates of the therapy have suggested that an explicit hospital‐based plan developed by key stakeholders can help facilitate implementation.21 Accordingly, there is growing interest in developing centers of expertise in highly intensive therapies such as TH. For instance, the New York City Emergency Medical Service has begun to explore a protocol to divert TH candidates to specialized centers.22, 23 Some favorable results have been reported in individual hospitals and local hospital systems.2428

Our data suggest that TH is associated with an in‐hospital mortality rate that is comparable to that reported in the clinical trials. For example, in a 2009 meta‐analysis of 4 clinical trials and 1 abstract (481 patients in total), TH was associated with a 35% relative mortality benefit as compared to standard post‐resuscitation care.29 It has been estimated that broad TH implementation could save thousands of lives30 and many authors have advocated for its use and outlined explicit protocols for implementation.17 Furthermore, TH appears to be cost‐effective in line with other accepted therapies. Assuming the Hypothermia After Cardiac Arrest (HACA) trial inclusion criteria, even at extreme estimates for costs, the cost‐effectiveness of hypothermia remains less than $100,000 per quality‐adjusted life year.31

There are important limitations of this study. Our use of administrative claims data certainly underestimates the level of TH utilization, since we could only identify cases in which TH was included in the billing codes for the hospitalization. Hospitals may vary in utilization of this particular billing code for TH in ways that bias our estimated associations. The ICD‐9 code 99.81 for therapeutic hypothermia was also not developed for post‐cardiac arrest TH specifically, so use of the code may actually lag clinical utilization. Although the observed trend in TH utilization is likely mainly due to a true increase in utilization, it is possible that some of the observed increase is due to an increase in utilization of TH procedure billing codes. Our TH utilization estimates should be construed as a lower bound of the actual rates. Additionally, although the estimated real‐world mortality benefit of TH may be comparable to that of clinical trials, the equivalence of patients in our sample to those in published randomized trials is uncertain. Similarly, even after adjusting for age, gender, year of admission, comorbidities, hospital teaching status, and TH volume, there are likely many unmeasured variables that influence mortality in both the TH and comparison groups. There are also likely patients included in our comparison group who had both cardiac arrest or ventricular tachycardia and anoxic brain injury, but who were not candidates for TH as the episode of cardiac arrest followed rather than preceded the anoxic brain injury. Since we lack detailed clinical data about the TH cases (initial rhythm, time before return of circulation, preexisting disease states, etc.), we are unable to match controls directly to cases. Additionally, we lack data to assess neurologic recovery or quality of life after arrest.

The observation that a mortality benefit in our study could be detected only in low and moderate volume centers requires further exploration. Indeed, one might expect that high volume centers may have better outcomes with TH as a result of more robust infrastructure, technical experience, and available resources. Our finding that mortality benefits of TH appear concentrated in centers with low to moderate volume of TH utilization suggest at least 1 of 2 possibilities. First, low and moderate volume centers may perform TH in a subset of patients who benefit most from the intervention or, alternatively, in the most viable cardiac arrest cases (those who may fare well with or without the therapy). Consequently, we may observe relatively favorable outcomes in this group due to this selection bias. Second, high volume centersdespite having more expertisemay also attract patients at higher mortality risk due to referral bias. This would lead us to estimate lower mortality benefits associated with TH in these high volume centers. Indeed, greater observed mortality at high volume centers regardless of TH status suggests that overall acuity is higher at high volume centers. While our inferences are greatly affected by issues of case selection and referral bias, it also important to consider the possibility that the estimated mortality benefit of TH in higher volume centers is lower because of the selection of patients who do not meet current guidelines for treatment with TH. Distinguishing whether the selection of patients undergoing TH at high volume centers is appropriate or inappropriate based on current guidelines is an important issue that merits further research with datasets with more refined patient clinical information.

In summary, therapeutic hypothermia utilization is low, but the rate of implementation has increased since the publication of the initial clinical trials in 2002. The bulk of TH utilization appears limited to a subset of high volume centers, and most centers in California appear to have not used the therapy. Real‐world in‐hospital mortality associated with TH is comparable to that reported in randomized clinical trials.

There are over 350,000 cases of out‐of‐hospital cardiac arrest (OHCA) each year in the United States1, 2 and, with supportive therapy alone, only a fraction of victims survive to hospital discharge. Rapid intervention including cardiopulmonary resuscitation in the moments following arrest is critical to minimizing neurologic injury, morbidity, and mortality. In 2002, two small randomized controlled trials showed a survival benefit of therapeutic hypothermia (TH) when provided within 12 hours after return of circulation following an OHCA and, to date, TH remains one of the few interventions with proven mortality benefit after initial cardiopulmonary resuscitation.3, 4 Since 2003, TH has been incorporated into the American Heart Association practice guidelines57 and use of TH has steadily increased, but widespread clinical uptake remains low.8, 9

The initial studies that evaluated TH were small, with only 189 patients included in the TH arms of the 2 trials combined. To date, only a few studies have replicated this initial observation in real‐world settings, with little analysis of outcomes in US centers in particular.1013 Accordingly, we aimed to examine the real‐world experience with TH in the United States using a large administrative claims database of all California hospital admissions to describe utilization trends, hospital mortality, and volumeoutcome relationships associated with the intervention.

MATERIALS AND METHODS

RESULTS

Descriptive Data

Table 1 reports summary statistics for patients with cardiac arrest complicated by neurologic insult (anoxic brain injury, coma, or persistent vegetative state) between 1999 and 2008. Across all years, 204 patients were identified as undergoing TH. In comparison, 105 patients were identified as undergoing TH in 2008 alone. Patients who underwent TH were less likely to be male (30.7% vs 44.6% male, P < 0.01), were younger (63.9 15.0 years vs 67.3 15.7 years, P = 0.03), and had equivalent numbers of Charlson‐Deyo comorbidities (2.5 2.0 diagnoses vs 2.5 2.0 diagnoses, P = 0.89). Therapeutic hypothermia was more commonly employed at teaching hospitals (51/3709 [1.4%] vs 153/42,942 [0.4%], P < 0.01). There was a trend toward decreased unadjusted mortality among patients who underwent therapeutic hypothermia compared with those who did not (56.9% vs 62.8%, P = 0.08).

Table 1.Characteristics of Patients Suffering Cardiac Arrest Complicated by Neurologic Insult, Stratified by Therapeutic Hypothermia Use, California Hospitals 19992008

	Therapeutic Hypothermia	No Therapeutic Hypothermia	P Value
P values for age, hospital mortality, and number of Charlson‐Deyo comorbidities reflect 2‐sided t test of continuous variables; P values for all other variables reflect 2 test of categorical data.
No. observations	204	46,629
No. cases in teaching hospitals	51	3,658
No. cases in non‐teaching hospitals	153	42,789
Age, y	63.9 15.0	67.3 15.7	0.06
Male	30.7	44.6	<0.01
Hospital mortality, %	56.9	62.8	0.08
Comorbidities
No. Charlson‐Deyo comorbidities	2.5 2.0	2.5 2.0	0.89
Coronary artery disease, %	48.0	38.0	<0.01
Acute myocardial infarction, %	42.6	28.9	<0.01
Congestive heart failure, %	27.9	35.3	0.03
Hypertension, %	36.3	33.2	0.83
Acute renal failure, %	33.3	26.6	0.03
Diabetes mellitus, %	30.9	23.0	<0.01
Chronic obstructive pulmonary disease, %	10.3	19.3	<0.01

Figures 1 and 2 provide additional aggregate statistics on therapeutic hypothermia in California hospitals. Figure 1 plots the number of therapeutic hypothermia cases recorded in the administrative discharge registry between 1999 and 2008. Of the 204 total cases identified during this period, 178 (87.3%) were performed between 2006 and 2008. Figure 2 shows the distribution of TH cases across centers that performed therapeutic hypothermia (n = 47 hospitals, 11.3 % of all hospitals). Ten centers accounted for 124/204 (60.7%) of the total patients treated with TH after cardiac arrest; the top 3 centers accounted for 64 (31.4%) of the treated patients. Twenty‐seven hospitals were identified as performing therapeutic hypothermia on only 1 or 2 patients between 1999 and 2008.

Figure 1

Figure 2

Risk‐Adjusted Mortality

Table 2 presents the odds ratio of factors predicting in‐hospital mortality after cardiac arrest complicated by neurologic insult. Factors include use of TH after cardiac arrest, age, gender, year of admission, number of Charlson‐Deyo comorbidities, hospital teaching status, and volume tercile of hospitals that performed therapeutic hypothermia. Overall, patients who were older, male, and had greater comorbidities were statistically more likely to die after cardiac arrest complicated by neurologic insult. Regardless of whether they underwent TH, patients admitted to hospitals in the highest volume tercile of TH use were more likely to die after cardiac arrest. Adjusting for volume tercile, teaching hospital status was not independently associated with mortality after cardiac arrest. The adjusted odds ratio of mortality among patients undergoing therapeutic hypothermia was 0.80 (95% confidence interval [CI] 0.601.06, P = 0.11). The adjusted probability of inpatient mortality among patients undergoing therapeutic hypothermia was 57.5% (95% CI 50.764.3%) compared to those who did not 62.8% (95% CI 61.763.9%, P = 0.11).

Table 2.Adjusted Odds Ratio of Hospital Mortality After Cardiac Arrest Complicated by Neurologic Insult in a Multivariable Regression Model, California Hospitals 19992008

Variable	Odds Ratio of Hospital Mortality (95% CI)	P Value
Abbreviations: CI, confidence interval; TH, therapeutic hypothermia. *Odds ratios compared to individuals aged 6065. Odds ratios compared to hospitals not performing TH.
No. observations	46,651
Age*
6569	1.19 (1.121.28)	<0.001
7074	1.29 (1.201.39)	<0.001
7579	1.55 (1.441.67)	<0.001
8084	1.79 (1.651.93)	<0.001
85 and over	2.06 (1.892.25)	<0.001
Male	1.15 (1.101.21)	<0.001
Teaching hospital	1.13 (0.951.34)	0.17
No. Charlson‐Deyo comorbidities	1.09 (1.081.10)	<0.001
Year trend	0.98 (0.970.99)	<0.001
Volume tercile among hospitals performing TH
First tercile	0.94 (0.791.12)	0.48
Second tercile	1.03 (0.801.33)	0.82
Third tercile	1.20 (1.051.36)	0.006
Therapeutic hypothermia	0.80 (0.601.06)	0.11

Figure 3 presents adjusted mortality after cardiac arrest in hospitals that did not perform TH, as well as adjusted mortality associated with TH for each volume tercile of hospitals that performed the procedure. Hospital mortality rates among patients not receiving TH after cardiac arrest were slightly higher in hospitals in the high volume tercile of TH (66.3%, 95% CI 63.868.8%) compared to hospitals in low and moderate volume terciles and to hospitals not performing TH (P < 0.001). Hospital mortality rates among low and moderate TH volume centers and in centers not performing TH were similar (62.3%, 61.3%, and 63.4%, respectively). Among both the low volume and moderate volume terciles, however, patients who underwent TH after cardiac arrest were significantly less likely to die in‐hospital compared to those who did not. For patients admitted to hospitals in the low volume tercile, those undergoing therapeutic hypothermia had an adjusted hospital mortality rate of 25.5% (95% CI 3.047.9%) compared to those who did not undergo TH (adjusted mortality 61.3%, 95% CI 57.465.1%), P < 0.001. In the moderate volume tercile, patients receiving therapeutic hypothermia had an adjusted hospital mortality rate of 31.0% (95% CI 9.2%52.8%) compared to 63.4% (95% CI 57.769.1%), P < 0.001, among those not undergoing the procedure. There was no statistically significant difference in adjusted mortality between those who underwent TH and those who did not, in hospitals in the highest volume tercile (P = 0.211). In addition to examining how volume of therapeutic hypothermia performed by hospitals affected the association between TH and hospital mortality, we also examined whether year of admission and teaching hospital independently modified the association. Neither year of admission nor teaching hospital statistically significantly affected the association between therapeutic hypothermia and hospital mortality after cardiac arrest at the P < 0.10 level.

Figure 3

DISCUSSION

In an administrative database of all admissions to California hospitals, we demonstrated that use of TH increased steadily since the publication of the initial clinical trials in 2002. The absolute level of TH utilization in our study undoubtedly represents a significant underestimation of actual TH utilization, however, our study does provide an assessment of the utilization trends over time. The bulk of TH use appears to be performed in a small group of high volume centers, and 89% of California hospitals did not perform TH during the study period (as assessed by procedure billing codes). Additionally, within the limitations of a retrospective, administrative claims‐based study design, TH appears to be associated with a similar in‐hospital mortality rate to that seen in clinical trials.3, 4 In exploratory analyses, there appears to be a particular benefit of TH in low and moderate volume centers, though these findings should be considered hypothesis‐generating.

Despite the body of evidence in favor of TH, utilization in our study and others appears quite low. In a 2005 survey of physicians, 87% of respondents had never used therapeutic hypothermia, citing inadequate data, technical limitations, and lack of incorporation in the Advanced Cardiac Life Support (ACLS) protocol as principal justifications.8 Other surveys have shown similar results and noted that critical care physicians and those working in large medical centers were more likely to adopt the therapy.9 Advocates of the therapy have suggested that an explicit hospital‐based plan developed by key stakeholders can help facilitate implementation.21 Accordingly, there is growing interest in developing centers of expertise in highly intensive therapies such as TH. For instance, the New York City Emergency Medical Service has begun to explore a protocol to divert TH candidates to specialized centers.22, 23 Some favorable results have been reported in individual hospitals and local hospital systems.2428

Our data suggest that TH is associated with an in‐hospital mortality rate that is comparable to that reported in the clinical trials. For example, in a 2009 meta‐analysis of 4 clinical trials and 1 abstract (481 patients in total), TH was associated with a 35% relative mortality benefit as compared to standard post‐resuscitation care.29 It has been estimated that broad TH implementation could save thousands of lives30 and many authors have advocated for its use and outlined explicit protocols for implementation.17 Furthermore, TH appears to be cost‐effective in line with other accepted therapies. Assuming the Hypothermia After Cardiac Arrest (HACA) trial inclusion criteria, even at extreme estimates for costs, the cost‐effectiveness of hypothermia remains less than $100,000 per quality‐adjusted life year.31

There are important limitations of this study. Our use of administrative claims data certainly underestimates the level of TH utilization, since we could only identify cases in which TH was included in the billing codes for the hospitalization. Hospitals may vary in utilization of this particular billing code for TH in ways that bias our estimated associations. The ICD‐9 code 99.81 for therapeutic hypothermia was also not developed for post‐cardiac arrest TH specifically, so use of the code may actually lag clinical utilization. Although the observed trend in TH utilization is likely mainly due to a true increase in utilization, it is possible that some of the observed increase is due to an increase in utilization of TH procedure billing codes. Our TH utilization estimates should be construed as a lower bound of the actual rates. Additionally, although the estimated real‐world mortality benefit of TH may be comparable to that of clinical trials, the equivalence of patients in our sample to those in published randomized trials is uncertain. Similarly, even after adjusting for age, gender, year of admission, comorbidities, hospital teaching status, and TH volume, there are likely many unmeasured variables that influence mortality in both the TH and comparison groups. There are also likely patients included in our comparison group who had both cardiac arrest or ventricular tachycardia and anoxic brain injury, but who were not candidates for TH as the episode of cardiac arrest followed rather than preceded the anoxic brain injury. Since we lack detailed clinical data about the TH cases (initial rhythm, time before return of circulation, preexisting disease states, etc.), we are unable to match controls directly to cases. Additionally, we lack data to assess neurologic recovery or quality of life after arrest.

The observation that a mortality benefit in our study could be detected only in low and moderate volume centers requires further exploration. Indeed, one might expect that high volume centers may have better outcomes with TH as a result of more robust infrastructure, technical experience, and available resources. Our finding that mortality benefits of TH appear concentrated in centers with low to moderate volume of TH utilization suggest at least 1 of 2 possibilities. First, low and moderate volume centers may perform TH in a subset of patients who benefit most from the intervention or, alternatively, in the most viable cardiac arrest cases (those who may fare well with or without the therapy). Consequently, we may observe relatively favorable outcomes in this group due to this selection bias. Second, high volume centersdespite having more expertisemay also attract patients at higher mortality risk due to referral bias. This would lead us to estimate lower mortality benefits associated with TH in these high volume centers. Indeed, greater observed mortality at high volume centers regardless of TH status suggests that overall acuity is higher at high volume centers. While our inferences are greatly affected by issues of case selection and referral bias, it also important to consider the possibility that the estimated mortality benefit of TH in higher volume centers is lower because of the selection of patients who do not meet current guidelines for treatment with TH. Distinguishing whether the selection of patients undergoing TH at high volume centers is appropriate or inappropriate based on current guidelines is an important issue that merits further research with datasets with more refined patient clinical information.

In summary, therapeutic hypothermia utilization is low, but the rate of implementation has increased since the publication of the initial clinical trials in 2002. The bulk of TH utilization appears limited to a subset of high volume centers, and most centers in California appear to have not used the therapy. Real‐world in‐hospital mortality associated with TH is comparable to that reported in randomized clinical trials.

Noncardiac surgery is frequently associated with major adverse cardiac events. Prevention of these events has traditionally focused on risk estimation and targeted interventions prior to surgical intervention.1 Cardiac risk is assessed through clinical prediction rules, such as the revised cardiac risk index (RCRI) or the American Society of Anesthesiology (ASA) classification. In patients deemed to be at high risk of adverse cardiac events, discretionary preoperative testing, medical treatments, and interventions are implemented.2 However, even when executed optimally, this approach fails to protect all patients. Thus, many patients undergoing noncardiac surgery continue to experience perioperative myocardial infarction (MI) and death.3

Recently, a prospective international study involving 15,133 patients reported that cardiac troponin levels measured within 3 days of noncardiac surgery were strongly associated with 30‐day mortality.4 This study is but the latest in a series of investigations that have concluded that postoperative measurement of cardiac troponin is a better predictor of cardiac outcomes than preoperative‐risk algorithms.48 Intuitively, this link is not surprising. Surgery represents the ultimate physiologic stress test. In the setting of hemodynamic changes and sustained oxygen supply:demand mismatch, it is hardly shocking that those with significant epicardial coronary stenosis or major subendocardial ischemia suffer poorer outcomes. What is perhaps most important about the association between troponin release and clinical outcomes, however, is the foundation it provides for a novel framework to improve perioperative care: using troponin measurement to guide treatment and interventions in the postoperative setting.

MODEL TO IMPROVE POSTOPERATIVE CARDIAC OUTCOMES

We hypothesize that the postoperative period can be seized as a golden moment to ameliorate cardiac risk and improve clinical outcomes for 3 reasons. First, those who release cardiac troponin during surgery declare themselves as harboring myocardial territories in jeopardy. In effect, the peripheral presence of troponin amounts to a cry for help in a population where many cardiac events are silent and thus remain clinically unnoticed.9, 10 Routine postoperative monitoring for this biomarker can thus help identify a population that would ordinarily escape detection until the precipitation of a major clinical event. Second, as the pathophysiology of cardiac events centers around rupture of vulnerable plaque (type‐1 MI), supply:demand mismatch (type‐2 MI), or both of these phenomena, escalation or institution of important medical risk‐reducing treatments (such as aspirin, beta‐blockers, or statins), can be established to improve outcomes. The ability to initiate these treatments in a monitored setting may offset and address many of the risks feared with these therapies, including hypotension, stroke, or rhabdomyolysis. Third, as patients convalesce from surgery, timely and discretionary cardiac testing or interventions can be undertaken in those identified as being at elevated cardiovascular risk prior to discharge. In sum, the trifecta of clearly recognizing those at risk, instituting interventions in controlled settings, and performing well‐timed interventions could translate into greater odds of survival in this cohort.

Owing to their intimate involvement with patients in the perioperative setting, hospitalists are uniquely suited for implementing a troponin surveillance paradigm. Yet, how may such a schema be realized? In Figure 1, we outline a pathway with which to implement this approach. This model replicates perioperative management workflow, calling for 3 discrete interventionspreoperative assessment, perioperative surveillance, and postoperative risk stratification:

Figure 1

• 1

  Preoperative assessment: Patients estimated to be at high risk of perioperative cardiac events (eg, RCRI 3) should undergo baseline, preoperative high‐sensitivity troponin measurement, as the frequency of major adverse cardiac outcomes is significant this group (11%).11 The importance of a baseline measurement of troponin cannot be overstated, as biologic variation of this marker can impact subsequent management.12, 13 Hospitalists can obtain these assays in preoperative clinics at the time of initial review and risk‐estimation.

• 2

  Perioperative surveillance: Following the model implemented by a recent study, serial troponin surveillance should commence 612 hours after surgery, then daily on the first, second, and third day.4 As the optimal cardiac troponin threshold for the diagnosis of perioperative ischemia remains uncertain, doubling of baseline troponin levels may serve as a logical signal for further evaluation.14 The electrocardiographic pattern may be used as a branch point for immediate management at this stage: ST‐segment elevation MI may be treated aggressively with an early invasive strategy and cardiac catheterization when apropos (much as in the nonoperative context). However, as most perioperative cardiac events are non‐ST elevation MIs, treatment of this subset should center on initiation or escalation of medical therapy. Thus, achieving heart‐rate control in order to preserve coronary perfusion by careful titration of beta‐blockers, and initiation/up‐titration of statins to temper unstable atherosclerotic plaques represent cornerstones of this approach. Other potential therapeutic entities such as initiation of angiotensin‐converting‐enzyme (ACE)‐inhibitors in the context of left‐ventricular dysfunction, and/or antiplatelet agents such as aspirin, may also be relevant risk‐reducing tools.

• 3

  Postoperative risk stratification: Prior to hospital discharge, objective assessment of myocardial perfusion and viability should occur in selected individuals with troponin elevations. Modalities that may be utilized in this context include either noninvasive imaging or occasional coronary angiography in patients with unstable coronary syndromes. Defining an optimal approach must be contextual, as interventions that may subsequently mandate uninterrupted antiplatelet treatment may be logistically challenging in this setting.

LIMITATIONS

Although this implementation model is plausible, it generates many questions that must be tested through the lens of a randomized controlled study. For instance, what is the optimal strategy for intervention in the postoperative setting? Though medical treatment with statins and beta‐blockade may represent the mainstay of treatment, are these therapies safe during potential clinical instability? Can net benefit be realized through judicious use of coronary angiography and revascularization? Can the artifact of heightened awareness and reporting of postoperative cardiac events mar the reported quality of a hospital? Finally, though cardiac troponin has been shown to be a strong and independent predictor of mortality, which troponin assay, reference range, and troponometric standard to use remain unclear.15 Whether or not these interventions translate into lesser mortality and improved clinical outcomes represents the raison d'etre for this experimental approach. As no alternative approach to define and target patients at high risk of adverse outcome after seemingly uneventful surgery currently exists, we believe that this hypothetical paradigm is worthy of further investigation.

CONCLUSION

For decades, perioperative practitioners have searched for a divining rod to reduce risk during surgery. While our efforts have been focused on the preoperative setting, the postoperative setting represents an as yet untapped and potentially profound period in the quest to improving surgical outcomes. Through our proximity to patients in the perioperative setting, hospitalists are ideal agents to test, deliver, and bring about this change. It is time for a postoperative carpe diem.

Noncardiac surgery is frequently associated with major adverse cardiac events. Prevention of these events has traditionally focused on risk estimation and targeted interventions prior to surgical intervention.1 Cardiac risk is assessed through clinical prediction rules, such as the revised cardiac risk index (RCRI) or the American Society of Anesthesiology (ASA) classification. In patients deemed to be at high risk of adverse cardiac events, discretionary preoperative testing, medical treatments, and interventions are implemented.2 However, even when executed optimally, this approach fails to protect all patients. Thus, many patients undergoing noncardiac surgery continue to experience perioperative myocardial infarction (MI) and death.3

Recently, a prospective international study involving 15,133 patients reported that cardiac troponin levels measured within 3 days of noncardiac surgery were strongly associated with 30‐day mortality.4 This study is but the latest in a series of investigations that have concluded that postoperative measurement of cardiac troponin is a better predictor of cardiac outcomes than preoperative‐risk algorithms.48 Intuitively, this link is not surprising. Surgery represents the ultimate physiologic stress test. In the setting of hemodynamic changes and sustained oxygen supply:demand mismatch, it is hardly shocking that those with significant epicardial coronary stenosis or major subendocardial ischemia suffer poorer outcomes. What is perhaps most important about the association between troponin release and clinical outcomes, however, is the foundation it provides for a novel framework to improve perioperative care: using troponin measurement to guide treatment and interventions in the postoperative setting.

MODEL TO IMPROVE POSTOPERATIVE CARDIAC OUTCOMES

We hypothesize that the postoperative period can be seized as a golden moment to ameliorate cardiac risk and improve clinical outcomes for 3 reasons. First, those who release cardiac troponin during surgery declare themselves as harboring myocardial territories in jeopardy. In effect, the peripheral presence of troponin amounts to a cry for help in a population where many cardiac events are silent and thus remain clinically unnoticed.9, 10 Routine postoperative monitoring for this biomarker can thus help identify a population that would ordinarily escape detection until the precipitation of a major clinical event. Second, as the pathophysiology of cardiac events centers around rupture of vulnerable plaque (type‐1 MI), supply:demand mismatch (type‐2 MI), or both of these phenomena, escalation or institution of important medical risk‐reducing treatments (such as aspirin, beta‐blockers, or statins), can be established to improve outcomes. The ability to initiate these treatments in a monitored setting may offset and address many of the risks feared with these therapies, including hypotension, stroke, or rhabdomyolysis. Third, as patients convalesce from surgery, timely and discretionary cardiac testing or interventions can be undertaken in those identified as being at elevated cardiovascular risk prior to discharge. In sum, the trifecta of clearly recognizing those at risk, instituting interventions in controlled settings, and performing well‐timed interventions could translate into greater odds of survival in this cohort.

Owing to their intimate involvement with patients in the perioperative setting, hospitalists are uniquely suited for implementing a troponin surveillance paradigm. Yet, how may such a schema be realized? In Figure 1, we outline a pathway with which to implement this approach. This model replicates perioperative management workflow, calling for 3 discrete interventionspreoperative assessment, perioperative surveillance, and postoperative risk stratification:

Figure 1

• 1

  Preoperative assessment: Patients estimated to be at high risk of perioperative cardiac events (eg, RCRI 3) should undergo baseline, preoperative high‐sensitivity troponin measurement, as the frequency of major adverse cardiac outcomes is significant this group (11%).11 The importance of a baseline measurement of troponin cannot be overstated, as biologic variation of this marker can impact subsequent management.12, 13 Hospitalists can obtain these assays in preoperative clinics at the time of initial review and risk‐estimation.

• 2

  Perioperative surveillance: Following the model implemented by a recent study, serial troponin surveillance should commence 612 hours after surgery, then daily on the first, second, and third day.4 As the optimal cardiac troponin threshold for the diagnosis of perioperative ischemia remains uncertain, doubling of baseline troponin levels may serve as a logical signal for further evaluation.14 The electrocardiographic pattern may be used as a branch point for immediate management at this stage: ST‐segment elevation MI may be treated aggressively with an early invasive strategy and cardiac catheterization when apropos (much as in the nonoperative context). However, as most perioperative cardiac events are non‐ST elevation MIs, treatment of this subset should center on initiation or escalation of medical therapy. Thus, achieving heart‐rate control in order to preserve coronary perfusion by careful titration of beta‐blockers, and initiation/up‐titration of statins to temper unstable atherosclerotic plaques represent cornerstones of this approach. Other potential therapeutic entities such as initiation of angiotensin‐converting‐enzyme (ACE)‐inhibitors in the context of left‐ventricular dysfunction, and/or antiplatelet agents such as aspirin, may also be relevant risk‐reducing tools.

• 3

  Postoperative risk stratification: Prior to hospital discharge, objective assessment of myocardial perfusion and viability should occur in selected individuals with troponin elevations. Modalities that may be utilized in this context include either noninvasive imaging or occasional coronary angiography in patients with unstable coronary syndromes. Defining an optimal approach must be contextual, as interventions that may subsequently mandate uninterrupted antiplatelet treatment may be logistically challenging in this setting.

LIMITATIONS

Although this implementation model is plausible, it generates many questions that must be tested through the lens of a randomized controlled study. For instance, what is the optimal strategy for intervention in the postoperative setting? Though medical treatment with statins and beta‐blockade may represent the mainstay of treatment, are these therapies safe during potential clinical instability? Can net benefit be realized through judicious use of coronary angiography and revascularization? Can the artifact of heightened awareness and reporting of postoperative cardiac events mar the reported quality of a hospital? Finally, though cardiac troponin has been shown to be a strong and independent predictor of mortality, which troponin assay, reference range, and troponometric standard to use remain unclear.15 Whether or not these interventions translate into lesser mortality and improved clinical outcomes represents the raison d'etre for this experimental approach. As no alternative approach to define and target patients at high risk of adverse outcome after seemingly uneventful surgery currently exists, we believe that this hypothetical paradigm is worthy of further investigation.

CONCLUSION

For decades, perioperative practitioners have searched for a divining rod to reduce risk during surgery. While our efforts have been focused on the preoperative setting, the postoperative setting represents an as yet untapped and potentially profound period in the quest to improving surgical outcomes. Through our proximity to patients in the perioperative setting, hospitalists are ideal agents to test, deliver, and bring about this change. It is time for a postoperative carpe diem.

Noncardiac surgery is frequently associated with major adverse cardiac events. Prevention of these events has traditionally focused on risk estimation and targeted interventions prior to surgical intervention.1 Cardiac risk is assessed through clinical prediction rules, such as the revised cardiac risk index (RCRI) or the American Society of Anesthesiology (ASA) classification. In patients deemed to be at high risk of adverse cardiac events, discretionary preoperative testing, medical treatments, and interventions are implemented.2 However, even when executed optimally, this approach fails to protect all patients. Thus, many patients undergoing noncardiac surgery continue to experience perioperative myocardial infarction (MI) and death.3

Recently, a prospective international study involving 15,133 patients reported that cardiac troponin levels measured within 3 days of noncardiac surgery were strongly associated with 30‐day mortality.4 This study is but the latest in a series of investigations that have concluded that postoperative measurement of cardiac troponin is a better predictor of cardiac outcomes than preoperative‐risk algorithms.48 Intuitively, this link is not surprising. Surgery represents the ultimate physiologic stress test. In the setting of hemodynamic changes and sustained oxygen supply:demand mismatch, it is hardly shocking that those with significant epicardial coronary stenosis or major subendocardial ischemia suffer poorer outcomes. What is perhaps most important about the association between troponin release and clinical outcomes, however, is the foundation it provides for a novel framework to improve perioperative care: using troponin measurement to guide treatment and interventions in the postoperative setting.

MODEL TO IMPROVE POSTOPERATIVE CARDIAC OUTCOMES

We hypothesize that the postoperative period can be seized as a golden moment to ameliorate cardiac risk and improve clinical outcomes for 3 reasons. First, those who release cardiac troponin during surgery declare themselves as harboring myocardial territories in jeopardy. In effect, the peripheral presence of troponin amounts to a cry for help in a population where many cardiac events are silent and thus remain clinically unnoticed.9, 10 Routine postoperative monitoring for this biomarker can thus help identify a population that would ordinarily escape detection until the precipitation of a major clinical event. Second, as the pathophysiology of cardiac events centers around rupture of vulnerable plaque (type‐1 MI), supply:demand mismatch (type‐2 MI), or both of these phenomena, escalation or institution of important medical risk‐reducing treatments (such as aspirin, beta‐blockers, or statins), can be established to improve outcomes. The ability to initiate these treatments in a monitored setting may offset and address many of the risks feared with these therapies, including hypotension, stroke, or rhabdomyolysis. Third, as patients convalesce from surgery, timely and discretionary cardiac testing or interventions can be undertaken in those identified as being at elevated cardiovascular risk prior to discharge. In sum, the trifecta of clearly recognizing those at risk, instituting interventions in controlled settings, and performing well‐timed interventions could translate into greater odds of survival in this cohort.

Owing to their intimate involvement with patients in the perioperative setting, hospitalists are uniquely suited for implementing a troponin surveillance paradigm. Yet, how may such a schema be realized? In Figure 1, we outline a pathway with which to implement this approach. This model replicates perioperative management workflow, calling for 3 discrete interventionspreoperative assessment, perioperative surveillance, and postoperative risk stratification:

Figure 1

• 1

  Preoperative assessment: Patients estimated to be at high risk of perioperative cardiac events (eg, RCRI 3) should undergo baseline, preoperative high‐sensitivity troponin measurement, as the frequency of major adverse cardiac outcomes is significant this group (11%).11 The importance of a baseline measurement of troponin cannot be overstated, as biologic variation of this marker can impact subsequent management.12, 13 Hospitalists can obtain these assays in preoperative clinics at the time of initial review and risk‐estimation.

• 2

  Perioperative surveillance: Following the model implemented by a recent study, serial troponin surveillance should commence 612 hours after surgery, then daily on the first, second, and third day.4 As the optimal cardiac troponin threshold for the diagnosis of perioperative ischemia remains uncertain, doubling of baseline troponin levels may serve as a logical signal for further evaluation.14 The electrocardiographic pattern may be used as a branch point for immediate management at this stage: ST‐segment elevation MI may be treated aggressively with an early invasive strategy and cardiac catheterization when apropos (much as in the nonoperative context). However, as most perioperative cardiac events are non‐ST elevation MIs, treatment of this subset should center on initiation or escalation of medical therapy. Thus, achieving heart‐rate control in order to preserve coronary perfusion by careful titration of beta‐blockers, and initiation/up‐titration of statins to temper unstable atherosclerotic plaques represent cornerstones of this approach. Other potential therapeutic entities such as initiation of angiotensin‐converting‐enzyme (ACE)‐inhibitors in the context of left‐ventricular dysfunction, and/or antiplatelet agents such as aspirin, may also be relevant risk‐reducing tools.

• 3

  Postoperative risk stratification: Prior to hospital discharge, objective assessment of myocardial perfusion and viability should occur in selected individuals with troponin elevations. Modalities that may be utilized in this context include either noninvasive imaging or occasional coronary angiography in patients with unstable coronary syndromes. Defining an optimal approach must be contextual, as interventions that may subsequently mandate uninterrupted antiplatelet treatment may be logistically challenging in this setting.

LIMITATIONS

Although this implementation model is plausible, it generates many questions that must be tested through the lens of a randomized controlled study. For instance, what is the optimal strategy for intervention in the postoperative setting? Though medical treatment with statins and beta‐blockade may represent the mainstay of treatment, are these therapies safe during potential clinical instability? Can net benefit be realized through judicious use of coronary angiography and revascularization? Can the artifact of heightened awareness and reporting of postoperative cardiac events mar the reported quality of a hospital? Finally, though cardiac troponin has been shown to be a strong and independent predictor of mortality, which troponin assay, reference range, and troponometric standard to use remain unclear.15 Whether or not these interventions translate into lesser mortality and improved clinical outcomes represents the raison d'etre for this experimental approach. As no alternative approach to define and target patients at high risk of adverse outcome after seemingly uneventful surgery currently exists, we believe that this hypothetical paradigm is worthy of further investigation.

CONCLUSION

For decades, perioperative practitioners have searched for a divining rod to reduce risk during surgery. While our efforts have been focused on the preoperative setting, the postoperative setting represents an as yet untapped and potentially profound period in the quest to improving surgical outcomes. Through our proximity to patients in the perioperative setting, hospitalists are ideal agents to test, deliver, and bring about this change. It is time for a postoperative carpe diem.

The sudden unexpected death of a hospitalized patient is extremely distressing to the family and the healthcare team. It is also distressingly common. Over 200,000 treated cardiac arrests are estimated to occur each year in US hospital patients.1 Most of these patients die, and studies have shown that physicians often incorrectly diagnose the causes of death when autopsies are performed to determine the causes of death.2, 3 This study was undertaken to elucidate the causes of sudden unexpected death of adult hospital patients as determined by autopsy.

METHODS

One hundred seventy‐five consecutive cases with autopsies by the senior author (L.N.) of adult hospital patients in the University of Pittsburgh Medical Center (UPMC) Health System, who died within 1 hour after onset of symptoms and in which death was unexpected, were retrospectively analyzed. Patients under 18, or dead on arrival, or on comfort measures only, were excluded. The unexpectedness of the deaths in this series was determined by review of the medical record, usually confirmed by a pre‐autopsy discussion with a clinician (following UPMC policy), and by the fact that attempted resuscitation was carried out in all but a few cases. Patient age, sex, race, and causes of death were obtained from the autopsy report. The medical record was reviewed to determine if the patient was on cardiac monitoring at the time of death. The study was approved by the University of Pittsburgh Medical Center Committee for Oversight of Research Involving the Dead.

RESULTS

The 175 autopsies in this study included 98 male patients and 77 female patients. Their ages ranged from 19 to 95 years, with an average age of 63.8 years. Categorized by race, 139 were white, 34 black, 1 Hispanic, and 1 Filipino. The autopsies were done over a 14‐year period from 1992 to 2006, during which the autopsy rate gradually decreased from 19% to 8%. Seeking authorization for autopsy from family was the responsibility of clinicians.

The most common immediate cause of death was judged to be a cardiac arrhythmia, usually presumptive; this was the immediate cause of death in 58 (33.1%) cases, as shown in Table 1. Second most common was hemorrhage, which was the immediate cause of death in 38 (21.7%) cases, and third was pulmonary thromboembolism in 27 (15.4%) cases, as shown in Table 1. Other conditions judged to be the immediate cause of death were cardiogenic pulmonary edema/congestive heart failure in 13 cases (7.4%), sepsis in 11 cases (6.3%), pulmonary edema due to acute lung injury in 4 cases (3 with associated pneumonia), and acute respiratory failure due to pneumonia (in 3 cases), usual interstitial pneumonia (2), emphysema (1), chronic obstructive pulmonary disease (1), herpes simplex virus bronchitis (1), carbon dioxide narcosis (1), and undiagnosed massive metastatic pancreatic carcinoma infiltrating and pushing up the diaphragm (1). Miscellaneous other conditions judged to be the immediate cause of death were brain stem compression in 2 cases, aspiration in 2 cases, andin 1 case eachsubarachnoid hemorrhage, hemorrhagic cerebral infarction, fat embolism, amniotic fluid embolism, bilateral pneumothoraces, massive hemolysis, sickle cell vaso‐occlusive crisis, cardiopulmonary decompensation, cardiac tamponade (due to pericardial metastases), and shock and systemic inflammatory response syndrome (due to volvulus).

The majority, 36 of the 58 patients (62%), with sudden death judged due to arrhythmias, had 75% or greater stenosis of 1 or more coronary arteries (average 2); 11 of these patients were on cardiac monitoring with their fatal arrhythmia displayed on the monitor, and 1 patient was wearing a Holter monitor at the time of her sudden death. The frequency of cardiac arrhythmia as a cause of sudden death did not change over the course of the 14 years of this study. Among the 31 of the 58 patients (53.4%) with histologically confirmed myocardial infarctions, 15 (25.9%) had a remote or subacute myocardial infarction without a history of myocardial infarction.

The most common underlying cause of death was severe coronary atherosclerosis, as shown in Table 2, but there were 14 patients whose underlying cause of death was a diverse group of other heart diseases. Five patients died postoperatively following heart surgery. One patient had a mitral valve papillary fibroelastoma, and another patient's arrhythmia was preceded by a right bundle branch block and a new first‐degree atrioventricular block attributable to mitral annular calcification. One patient with sickle cell disease had a myocardial bridge over a coronary artery. Three patients had heart transplants, 2 alcoholic cardiomyopathy, and 1 idiopathic dilated cardiomyopathy.

The initial cardiac rhythm during attempted cardiopulmonary resuscitation was obtainable for 120 cases. As shown in Table 3, a higher proportion of patients judged to have died of arrhythmias had an initial rhythm of ventricular tachycardia or fibrillation than the proportion of those judged to have died of hemorrhage, pulmonary embolism, or other immediate causes of death.

DISCUSSION

The patients who had an initial cardiac rhythm of asystole during attempted resuscitation, and were judged to have died of an arrhythmia, most likely had ventricular tachycardia or some other arrhythmia before asystole. The majority of them were found in cardiac arrest at night, not on cardiac monitoring. In other studies, there has been an unexplained continuing decline in the prevalence of sudden cardiac arrest cases presenting with ventricular fibrillation and corresponding rise in the prevalence of pulseless electrical activity (PEA) arrests.4 The 7 patients who had an initial cardiac rhythm of PEA and were judged to have died of an arrhythmia were all, except 1, off cardiac monitoring. Four had myocardial infarctions, 3 heart block pre‐arrest, 1 postoperative intramyocardial hematoma, 1 myocardial metastatic melanoma, 1 acute heart transplant rejection, and 3 ventricular tachycardia or fibrillation following PEA, suggesting that PEA was a phase in the evolution of an arrhythmogenic death. The patients who had an initial cardiac rhythm of ventricular tachycardia or fibrillation, and were judged to have died of hemorrhage or pulmonary thromboembolism, generally had coronary artery disease. The hemorrhage or embolism presumably caused terminal myocardial ischemia, but the cardiac disease was judged to be a contributing cause of death, less important than the immediate cause.

The conclusion that a cardiac arrhythmia was the most common mechanism of death in this series fits with the conclusion that cardiac arrhythmias were the most common immediate cause of cardiac arrest, specifically of 49% of them, in a study of 14,720 cardiac arrests of adult inpatients.5 It is accepted by convention that the presence of 75% or greater cross‐sectional luminal narrowing of a coronary artery, even without thrombosis, can be a cause of sudden cardiac death.6 Cardiac arrhythmias commonly occur during myocardial ischemia prior to the irreversible necrosis of myocardial infarction.7 Fatal cardiac arrhythmias can also be caused by old myocardial infarctions.8 The finding that 15 of the patients in this study, 25.9% of the 58 whose sudden death was attributable to a cardiac arrhythmia, had a remote or subacute myocardial infarction without a history of myocardial infarction fits with evidence that 25%30% of myocardial infarctions are unrecognized.

Hemorrhage has been an underpublicized cause of unexpected sudden death in hospital patients. Intracranial hemorrhages, due to ruptured berry aneurysms, hypertension, tumors, or arteriovenous malformations, are well‐recognized causes of sudden death, but not specifically in hospital patients.9 In a Scottish series of 111 unexpected sudden deaths due to an acute abdomen in patients aged 70 or older, 24 died of acute gastrointestinal hemorrhage, but cases of ruptured abdominal aortic aneurysm were excluded from the study because they would have dominated the analysis.10 Retroperitoneal hemorrhage is a particularly insidious cause of sudden death. It commonly causes little or no pain, and proceeds asymptomatically until the patient reaches the limits of cardiopulmonary compensation, which can mask the hemodynamic effect of the bleeding until sudden death.

The limitations of this study include the presumptive nature of the arrhythmias in the majority of patients judged most likely to have died of arrhythmias, and a potential selection bias in the cases coming to autopsy. No data on the 81% to 92% of deaths not investigated by autopsy is available, so the possibility of some sort of selection bias in this case series cannot be excluded. The causes of death determined by autopsy also inevitably represent a judgment or opinion about causation (as opposed to mere correlation), just as the assessment of the causes of death without autopsy does, but autopsy adds the knowledge of conditions undiagnosed prior to death and the exclusion of some suspected diagnoses, substantially improving the unavoidably judgmental conclusion.

There are implications for preventing the sudden unexpected death of hospital patients from the results of this study. They suggest that more cardiac rhythm monitoring might be helpful.11 More prophylactic antiarrhythmic medication and automatic implanted cardiac defibrillators might also be helpful.12 Some UPMC intensivists believe that the sort of fatal arrhythmias seen in this study are caused by hypoxemia, which suggests that more pulse oximetry oxygen saturation monitoring might allow preventative intervention.13 More frequent and possibly automated monitoring of vital signs might provide early warning of hemorrhage or pulmonary embolization.14 Keeping a wide differential diagnosis is taught for resuscitation with an initial rhythm of PEA, but keeping a wide differential including hemorrhage in cases with other initial rhythms, especially bradycardia, for example, may be important. This study suggests the importance of keeping hemorrhage in the differential diagnosis of sudden unexpected cardiac arrest of hospital patients.

The sudden unexpected death of a hospitalized patient is extremely distressing to the family and the healthcare team. It is also distressingly common. Over 200,000 treated cardiac arrests are estimated to occur each year in US hospital patients.1 Most of these patients die, and studies have shown that physicians often incorrectly diagnose the causes of death when autopsies are performed to determine the causes of death.2, 3 This study was undertaken to elucidate the causes of sudden unexpected death of adult hospital patients as determined by autopsy.

METHODS

One hundred seventy‐five consecutive cases with autopsies by the senior author (L.N.) of adult hospital patients in the University of Pittsburgh Medical Center (UPMC) Health System, who died within 1 hour after onset of symptoms and in which death was unexpected, were retrospectively analyzed. Patients under 18, or dead on arrival, or on comfort measures only, were excluded. The unexpectedness of the deaths in this series was determined by review of the medical record, usually confirmed by a pre‐autopsy discussion with a clinician (following UPMC policy), and by the fact that attempted resuscitation was carried out in all but a few cases. Patient age, sex, race, and causes of death were obtained from the autopsy report. The medical record was reviewed to determine if the patient was on cardiac monitoring at the time of death. The study was approved by the University of Pittsburgh Medical Center Committee for Oversight of Research Involving the Dead.

RESULTS

The 175 autopsies in this study included 98 male patients and 77 female patients. Their ages ranged from 19 to 95 years, with an average age of 63.8 years. Categorized by race, 139 were white, 34 black, 1 Hispanic, and 1 Filipino. The autopsies were done over a 14‐year period from 1992 to 2006, during which the autopsy rate gradually decreased from 19% to 8%. Seeking authorization for autopsy from family was the responsibility of clinicians.

The most common immediate cause of death was judged to be a cardiac arrhythmia, usually presumptive; this was the immediate cause of death in 58 (33.1%) cases, as shown in Table 1. Second most common was hemorrhage, which was the immediate cause of death in 38 (21.7%) cases, and third was pulmonary thromboembolism in 27 (15.4%) cases, as shown in Table 1. Other conditions judged to be the immediate cause of death were cardiogenic pulmonary edema/congestive heart failure in 13 cases (7.4%), sepsis in 11 cases (6.3%), pulmonary edema due to acute lung injury in 4 cases (3 with associated pneumonia), and acute respiratory failure due to pneumonia (in 3 cases), usual interstitial pneumonia (2), emphysema (1), chronic obstructive pulmonary disease (1), herpes simplex virus bronchitis (1), carbon dioxide narcosis (1), and undiagnosed massive metastatic pancreatic carcinoma infiltrating and pushing up the diaphragm (1). Miscellaneous other conditions judged to be the immediate cause of death were brain stem compression in 2 cases, aspiration in 2 cases, andin 1 case eachsubarachnoid hemorrhage, hemorrhagic cerebral infarction, fat embolism, amniotic fluid embolism, bilateral pneumothoraces, massive hemolysis, sickle cell vaso‐occlusive crisis, cardiopulmonary decompensation, cardiac tamponade (due to pericardial metastases), and shock and systemic inflammatory response syndrome (due to volvulus).

The majority, 36 of the 58 patients (62%), with sudden death judged due to arrhythmias, had 75% or greater stenosis of 1 or more coronary arteries (average 2); 11 of these patients were on cardiac monitoring with their fatal arrhythmia displayed on the monitor, and 1 patient was wearing a Holter monitor at the time of her sudden death. The frequency of cardiac arrhythmia as a cause of sudden death did not change over the course of the 14 years of this study. Among the 31 of the 58 patients (53.4%) with histologically confirmed myocardial infarctions, 15 (25.9%) had a remote or subacute myocardial infarction without a history of myocardial infarction.

The most common underlying cause of death was severe coronary atherosclerosis, as shown in Table 2, but there were 14 patients whose underlying cause of death was a diverse group of other heart diseases. Five patients died postoperatively following heart surgery. One patient had a mitral valve papillary fibroelastoma, and another patient's arrhythmia was preceded by a right bundle branch block and a new first‐degree atrioventricular block attributable to mitral annular calcification. One patient with sickle cell disease had a myocardial bridge over a coronary artery. Three patients had heart transplants, 2 alcoholic cardiomyopathy, and 1 idiopathic dilated cardiomyopathy.

The initial cardiac rhythm during attempted cardiopulmonary resuscitation was obtainable for 120 cases. As shown in Table 3, a higher proportion of patients judged to have died of arrhythmias had an initial rhythm of ventricular tachycardia or fibrillation than the proportion of those judged to have died of hemorrhage, pulmonary embolism, or other immediate causes of death.

DISCUSSION

The patients who had an initial cardiac rhythm of asystole during attempted resuscitation, and were judged to have died of an arrhythmia, most likely had ventricular tachycardia or some other arrhythmia before asystole. The majority of them were found in cardiac arrest at night, not on cardiac monitoring. In other studies, there has been an unexplained continuing decline in the prevalence of sudden cardiac arrest cases presenting with ventricular fibrillation and corresponding rise in the prevalence of pulseless electrical activity (PEA) arrests.4 The 7 patients who had an initial cardiac rhythm of PEA and were judged to have died of an arrhythmia were all, except 1, off cardiac monitoring. Four had myocardial infarctions, 3 heart block pre‐arrest, 1 postoperative intramyocardial hematoma, 1 myocardial metastatic melanoma, 1 acute heart transplant rejection, and 3 ventricular tachycardia or fibrillation following PEA, suggesting that PEA was a phase in the evolution of an arrhythmogenic death. The patients who had an initial cardiac rhythm of ventricular tachycardia or fibrillation, and were judged to have died of hemorrhage or pulmonary thromboembolism, generally had coronary artery disease. The hemorrhage or embolism presumably caused terminal myocardial ischemia, but the cardiac disease was judged to be a contributing cause of death, less important than the immediate cause.

The conclusion that a cardiac arrhythmia was the most common mechanism of death in this series fits with the conclusion that cardiac arrhythmias were the most common immediate cause of cardiac arrest, specifically of 49% of them, in a study of 14,720 cardiac arrests of adult inpatients.5 It is accepted by convention that the presence of 75% or greater cross‐sectional luminal narrowing of a coronary artery, even without thrombosis, can be a cause of sudden cardiac death.6 Cardiac arrhythmias commonly occur during myocardial ischemia prior to the irreversible necrosis of myocardial infarction.7 Fatal cardiac arrhythmias can also be caused by old myocardial infarctions.8 The finding that 15 of the patients in this study, 25.9% of the 58 whose sudden death was attributable to a cardiac arrhythmia, had a remote or subacute myocardial infarction without a history of myocardial infarction fits with evidence that 25%30% of myocardial infarctions are unrecognized.

Hemorrhage has been an underpublicized cause of unexpected sudden death in hospital patients. Intracranial hemorrhages, due to ruptured berry aneurysms, hypertension, tumors, or arteriovenous malformations, are well‐recognized causes of sudden death, but not specifically in hospital patients.9 In a Scottish series of 111 unexpected sudden deaths due to an acute abdomen in patients aged 70 or older, 24 died of acute gastrointestinal hemorrhage, but cases of ruptured abdominal aortic aneurysm were excluded from the study because they would have dominated the analysis.10 Retroperitoneal hemorrhage is a particularly insidious cause of sudden death. It commonly causes little or no pain, and proceeds asymptomatically until the patient reaches the limits of cardiopulmonary compensation, which can mask the hemodynamic effect of the bleeding until sudden death.

The limitations of this study include the presumptive nature of the arrhythmias in the majority of patients judged most likely to have died of arrhythmias, and a potential selection bias in the cases coming to autopsy. No data on the 81% to 92% of deaths not investigated by autopsy is available, so the possibility of some sort of selection bias in this case series cannot be excluded. The causes of death determined by autopsy also inevitably represent a judgment or opinion about causation (as opposed to mere correlation), just as the assessment of the causes of death without autopsy does, but autopsy adds the knowledge of conditions undiagnosed prior to death and the exclusion of some suspected diagnoses, substantially improving the unavoidably judgmental conclusion.

There are implications for preventing the sudden unexpected death of hospital patients from the results of this study. They suggest that more cardiac rhythm monitoring might be helpful.11 More prophylactic antiarrhythmic medication and automatic implanted cardiac defibrillators might also be helpful.12 Some UPMC intensivists believe that the sort of fatal arrhythmias seen in this study are caused by hypoxemia, which suggests that more pulse oximetry oxygen saturation monitoring might allow preventative intervention.13 More frequent and possibly automated monitoring of vital signs might provide early warning of hemorrhage or pulmonary embolization.14 Keeping a wide differential diagnosis is taught for resuscitation with an initial rhythm of PEA, but keeping a wide differential including hemorrhage in cases with other initial rhythms, especially bradycardia, for example, may be important. This study suggests the importance of keeping hemorrhage in the differential diagnosis of sudden unexpected cardiac arrest of hospital patients.

The sudden unexpected death of a hospitalized patient is extremely distressing to the family and the healthcare team. It is also distressingly common. Over 200,000 treated cardiac arrests are estimated to occur each year in US hospital patients.1 Most of these patients die, and studies have shown that physicians often incorrectly diagnose the causes of death when autopsies are performed to determine the causes of death.2, 3 This study was undertaken to elucidate the causes of sudden unexpected death of adult hospital patients as determined by autopsy.

METHODS

One hundred seventy‐five consecutive cases with autopsies by the senior author (L.N.) of adult hospital patients in the University of Pittsburgh Medical Center (UPMC) Health System, who died within 1 hour after onset of symptoms and in which death was unexpected, were retrospectively analyzed. Patients under 18, or dead on arrival, or on comfort measures only, were excluded. The unexpectedness of the deaths in this series was determined by review of the medical record, usually confirmed by a pre‐autopsy discussion with a clinician (following UPMC policy), and by the fact that attempted resuscitation was carried out in all but a few cases. Patient age, sex, race, and causes of death were obtained from the autopsy report. The medical record was reviewed to determine if the patient was on cardiac monitoring at the time of death. The study was approved by the University of Pittsburgh Medical Center Committee for Oversight of Research Involving the Dead.

RESULTS

The 175 autopsies in this study included 98 male patients and 77 female patients. Their ages ranged from 19 to 95 years, with an average age of 63.8 years. Categorized by race, 139 were white, 34 black, 1 Hispanic, and 1 Filipino. The autopsies were done over a 14‐year period from 1992 to 2006, during which the autopsy rate gradually decreased from 19% to 8%. Seeking authorization for autopsy from family was the responsibility of clinicians.

The most common immediate cause of death was judged to be a cardiac arrhythmia, usually presumptive; this was the immediate cause of death in 58 (33.1%) cases, as shown in Table 1. Second most common was hemorrhage, which was the immediate cause of death in 38 (21.7%) cases, and third was pulmonary thromboembolism in 27 (15.4%) cases, as shown in Table 1. Other conditions judged to be the immediate cause of death were cardiogenic pulmonary edema/congestive heart failure in 13 cases (7.4%), sepsis in 11 cases (6.3%), pulmonary edema due to acute lung injury in 4 cases (3 with associated pneumonia), and acute respiratory failure due to pneumonia (in 3 cases), usual interstitial pneumonia (2), emphysema (1), chronic obstructive pulmonary disease (1), herpes simplex virus bronchitis (1), carbon dioxide narcosis (1), and undiagnosed massive metastatic pancreatic carcinoma infiltrating and pushing up the diaphragm (1). Miscellaneous other conditions judged to be the immediate cause of death were brain stem compression in 2 cases, aspiration in 2 cases, andin 1 case eachsubarachnoid hemorrhage, hemorrhagic cerebral infarction, fat embolism, amniotic fluid embolism, bilateral pneumothoraces, massive hemolysis, sickle cell vaso‐occlusive crisis, cardiopulmonary decompensation, cardiac tamponade (due to pericardial metastases), and shock and systemic inflammatory response syndrome (due to volvulus).

The majority, 36 of the 58 patients (62%), with sudden death judged due to arrhythmias, had 75% or greater stenosis of 1 or more coronary arteries (average 2); 11 of these patients were on cardiac monitoring with their fatal arrhythmia displayed on the monitor, and 1 patient was wearing a Holter monitor at the time of her sudden death. The frequency of cardiac arrhythmia as a cause of sudden death did not change over the course of the 14 years of this study. Among the 31 of the 58 patients (53.4%) with histologically confirmed myocardial infarctions, 15 (25.9%) had a remote or subacute myocardial infarction without a history of myocardial infarction.

The most common underlying cause of death was severe coronary atherosclerosis, as shown in Table 2, but there were 14 patients whose underlying cause of death was a diverse group of other heart diseases. Five patients died postoperatively following heart surgery. One patient had a mitral valve papillary fibroelastoma, and another patient's arrhythmia was preceded by a right bundle branch block and a new first‐degree atrioventricular block attributable to mitral annular calcification. One patient with sickle cell disease had a myocardial bridge over a coronary artery. Three patients had heart transplants, 2 alcoholic cardiomyopathy, and 1 idiopathic dilated cardiomyopathy.

The initial cardiac rhythm during attempted cardiopulmonary resuscitation was obtainable for 120 cases. As shown in Table 3, a higher proportion of patients judged to have died of arrhythmias had an initial rhythm of ventricular tachycardia or fibrillation than the proportion of those judged to have died of hemorrhage, pulmonary embolism, or other immediate causes of death.

DISCUSSION

The patients who had an initial cardiac rhythm of asystole during attempted resuscitation, and were judged to have died of an arrhythmia, most likely had ventricular tachycardia or some other arrhythmia before asystole. The majority of them were found in cardiac arrest at night, not on cardiac monitoring. In other studies, there has been an unexplained continuing decline in the prevalence of sudden cardiac arrest cases presenting with ventricular fibrillation and corresponding rise in the prevalence of pulseless electrical activity (PEA) arrests.4 The 7 patients who had an initial cardiac rhythm of PEA and were judged to have died of an arrhythmia were all, except 1, off cardiac monitoring. Four had myocardial infarctions, 3 heart block pre‐arrest, 1 postoperative intramyocardial hematoma, 1 myocardial metastatic melanoma, 1 acute heart transplant rejection, and 3 ventricular tachycardia or fibrillation following PEA, suggesting that PEA was a phase in the evolution of an arrhythmogenic death. The patients who had an initial cardiac rhythm of ventricular tachycardia or fibrillation, and were judged to have died of hemorrhage or pulmonary thromboembolism, generally had coronary artery disease. The hemorrhage or embolism presumably caused terminal myocardial ischemia, but the cardiac disease was judged to be a contributing cause of death, less important than the immediate cause.

The conclusion that a cardiac arrhythmia was the most common mechanism of death in this series fits with the conclusion that cardiac arrhythmias were the most common immediate cause of cardiac arrest, specifically of 49% of them, in a study of 14,720 cardiac arrests of adult inpatients.5 It is accepted by convention that the presence of 75% or greater cross‐sectional luminal narrowing of a coronary artery, even without thrombosis, can be a cause of sudden cardiac death.6 Cardiac arrhythmias commonly occur during myocardial ischemia prior to the irreversible necrosis of myocardial infarction.7 Fatal cardiac arrhythmias can also be caused by old myocardial infarctions.8 The finding that 15 of the patients in this study, 25.9% of the 58 whose sudden death was attributable to a cardiac arrhythmia, had a remote or subacute myocardial infarction without a history of myocardial infarction fits with evidence that 25%30% of myocardial infarctions are unrecognized.

Hemorrhage has been an underpublicized cause of unexpected sudden death in hospital patients. Intracranial hemorrhages, due to ruptured berry aneurysms, hypertension, tumors, or arteriovenous malformations, are well‐recognized causes of sudden death, but not specifically in hospital patients.9 In a Scottish series of 111 unexpected sudden deaths due to an acute abdomen in patients aged 70 or older, 24 died of acute gastrointestinal hemorrhage, but cases of ruptured abdominal aortic aneurysm were excluded from the study because they would have dominated the analysis.10 Retroperitoneal hemorrhage is a particularly insidious cause of sudden death. It commonly causes little or no pain, and proceeds asymptomatically until the patient reaches the limits of cardiopulmonary compensation, which can mask the hemodynamic effect of the bleeding until sudden death.

The limitations of this study include the presumptive nature of the arrhythmias in the majority of patients judged most likely to have died of arrhythmias, and a potential selection bias in the cases coming to autopsy. No data on the 81% to 92% of deaths not investigated by autopsy is available, so the possibility of some sort of selection bias in this case series cannot be excluded. The causes of death determined by autopsy also inevitably represent a judgment or opinion about causation (as opposed to mere correlation), just as the assessment of the causes of death without autopsy does, but autopsy adds the knowledge of conditions undiagnosed prior to death and the exclusion of some suspected diagnoses, substantially improving the unavoidably judgmental conclusion.

There are implications for preventing the sudden unexpected death of hospital patients from the results of this study. They suggest that more cardiac rhythm monitoring might be helpful.11 More prophylactic antiarrhythmic medication and automatic implanted cardiac defibrillators might also be helpful.12 Some UPMC intensivists believe that the sort of fatal arrhythmias seen in this study are caused by hypoxemia, which suggests that more pulse oximetry oxygen saturation monitoring might allow preventative intervention.13 More frequent and possibly automated monitoring of vital signs might provide early warning of hemorrhage or pulmonary embolization.14 Keeping a wide differential diagnosis is taught for resuscitation with an initial rhythm of PEA, but keeping a wide differential including hemorrhage in cases with other initial rhythms, especially bradycardia, for example, may be important. This study suggests the importance of keeping hemorrhage in the differential diagnosis of sudden unexpected cardiac arrest of hospital patients.

Fever is a nonspecific phenomenon that can result from many inciting causes such as infection, inflammation, malignancy, thromboembolic disease, drugs, and endocrine disease. In hospitalized patients, one of the most important clinical considerations is bacteremia. Although vital signs compose 3 of the 4 current criteria for the diagnosis of Systemic Inflammatory Response Syndrome (SIRS),1, 2 they contribute little to the diagnosis of the cause, which can be inflammation or infection. Unfortunately, the physician's clinical diagnosis of bacteremia lacks both sensitivity and specificity.35 Blood culture acquisition is a simple and basic diagnostic procedure routinely used in clinical practice that yields essential information for the evaluation of various infectious diseases.6 Positive blood cultures can demonstrate not only an infectious cause of disease but also the microbiological response to antibiotic therapy.7 However, studies have reported that 35% to 50% of positive blood cultures are falsely positive owing to contamination.711 False‐positive cultures may lead to the use of inappropriate or unnecessary antibiotics, additional testing and consultation, and prolonged hospitalizations that increase patient care costs.9, 12

Nursing staff caring for patients are generally able to assess oral intake, general clinical state, and care requirements. Moreover, the nursing staff are often able to identify problems with patients before physicians.13 In Japan, nurse‐assessed food consumption of every meal is standardized, and is frequently regarded as an indicator of the patient's clinical status, akin to a vital sign. In this context, we hypothesized that quantitative variations in food consumption could accurately distinguish those patients with or without bacteremia.

MATERIALS AND METHODS

RESULTS

During the study period, 851 patients aged 16 to 99 years (66.8 16.6), were eligible for inclusion (Figure 1). Baseline characteristics of the subjects are given in Table 1. The mean body temperature ( standard deviation [SD]) was 38.1 1.1C, and the mean CRP level was 8.7 8.1 mg/dL. The results show that the patients had at least 2 SIRS criteria with elevations in temperature and heart rate. Of the 851 patients entered into the study, only 122 (14.3%) had positive blood cultures. Of these, 75 patients (8.8%) were considered to have true‐positive blood culture. In this study, blood cultures were taken at the time of onset of fever, whether that was a new inpatient admission, or during the course of an admission to the hospital. On average, blood cultures were drawn 12 days after admission (SD, 5.6 days). Despite the variation in onset of fever, the inverse relationship of blood culture positivity to decreased food consumption held true (data not shown). Gram‐positive and Gram‐negative organisms were obtained in near equal amounts. The main pathogens recovered from the true‐positive blood cultures were Gram‐positive cocci (26 patients [34.7% in true‐positive blood cultures]), and Gram‐negative bacilli (46 patients [61.3%]), as shown in Table 1. The underlying clinical diagnosis included 28 urinary tract infections; 9 catheter‐associated infections; 5 cases each of pneumonia and abscess; 3 cases of phlebitis; 2 cases of meningitis and osteomyelitis; 1 case each of infective endocarditis, decubitus ulcer, and pelvic infection; and 17 cases of infection with an unknown focus.

Figure 1

Low, moderate, and high food consumption groups consisted of 344 patients (40.4%), 152 patients (17.9%), and 354 patients (41.7%), respectively (Table 1). Of these, 63 patients, 6 patients, and 6 patients had bacteremia in the low, moderate, and high food consumption group, respectively. In order to distinguish those patients who had decreased food consumption compared to almost normal food consumption, low and moderate food consumption groups were combined and compared to the high food consumption group. Comparison of the combined low and moderate food consumption group versus the high food consumption group revealed a sensitivity of 92.0% and a negative predictive value of 98.3% for excluding true bacteremia. Conversely, the specificity (45.1%) and the positive predictive value (13.9%) were poor.

In the univariate analysis, the following variables were not statistically significantly associated with true bacteremia: age, heart rate, and leukocyte counts. Significant univariate predictors of bacteremia and their associated cut‐off points were body temperature of 36 or 38C (odds ratio [OR], 2.5; 95% CI, 1.54.4), CRP 10.0 mg/dL (OR, 2.0; 95% CI, 1.23.2), and food consumption (OR, 8.5; 95% CI, 3.818.6) (Table 2). There was no evidence of colinearity. In the final stepwise logistic regression (Table 3), the significant predictors of bacteremia were body temperature of 36 or 38C (OR, 2.4; 95% CI, 1.44.2; P = 0.002), C‐reactive protein of 10.0 mg/dL (OR, 1.9; 95% CI, 1.23.0; P = 0.011), and food consumption (OR, 7.5; 95% CI, 3.416.6; P < 0.001). We identified only 6 patients with bacteremia in the high food consumption group. Three of the patients had been previously treated with antibiotics for conditions including infective endocarditis, osteomyelitis, and myelodysplasic syndrome.

On further analysis, we excluded patients who had received antibiotics before blood culture acquisition. There were 661 patients in this subanalysis. Low, moderate, and high food consumption groups consisted of 282 patients (41.4%), 118 patients (17.3%), and 261 patients (38.3%), respectively. Of these, 50 patients (17.7%), 5 patients (4.2%), and 4 patients (1.5%) had bacteremia in the low, moderate, and high food consumption groups, respectively. The sensitivity and negative predictive values were 93.2% and 98.5%, respectively. In the stepwise logistic regression, significant predictors of bacteremia were body temperature of 36 or 38C (OR, 3.0; 95% CI, 1.55.6; P = 0.001), CRP 10.0 mg/dL (OR, 2.1; 95% CI, 1.23.7; P = 0.006), and food consumption (OR, 9.3; 95% CI, 3.326.1; P < 0.001).

DISCUSSION

We found that in a group of 851 Japanese patients who were suspected with bacterial infection, the estimated food consumption was negatively associated, both significantly and independently, with the subsequent isolation of microorganisms from their blood cultures. If validated in other studies, this simple rule of thumb can provide the clinician with reasonable confidence that a febrile patient has a low probability of being bacteremic, as long as the appetite remains normal. Both the sensitivity and the negative predictive value were extremely high at 92.3% and 98.3%, respectively, suggesting that adequate oral intake is a strong marker against the presence of bacteremia. In this study, it was the strongest predictor of bacteremia in multivariate analysis. After including only antibiotic‐naive patients, the sensitivity and the negative predictive values were 93.2% and 98.5%, respectively. Administration of antibiotics may lead to improved appetite in febrile patients despite bacteremia in the presence of fever, and therefore, inquiring about recent or current antimicrobial usage should be a requirement when considering oral intake as an indicator of bacteremia.

Our study has limitations. Since we did not make treatment decisions based on oral intake, we cannot conclude that it is safe to withhold antibiotic treatment on the basis of food intake alone. Additionally, this study would need to be repeated across many different age groups and racial groups to ensure applicability to the general population. It is also unknown whether this rule would be applicable to patients with underlying immunosuppression. Finally, although inter‐rater reliability was high in our center, nurses in other settings may not be as diligent in their assessment of food consumption. The high inter‐assessor reliability in our setting, however, suggests that objective assessment of food intake can be performed reliably in settings in which accurate documentation of food consumption is expected.

In summary, we found that normal food intake was strongly and negatively associated with bacteremia in febrile patients. This observation, if validated in other settings, may serve as a simple aid to assist in the clinical diagnosis of bacteremia or for recruitment of patients with a high likelihood of bacteremia into clinical trials.

Fever is a nonspecific phenomenon that can result from many inciting causes such as infection, inflammation, malignancy, thromboembolic disease, drugs, and endocrine disease. In hospitalized patients, one of the most important clinical considerations is bacteremia. Although vital signs compose 3 of the 4 current criteria for the diagnosis of Systemic Inflammatory Response Syndrome (SIRS),1, 2 they contribute little to the diagnosis of the cause, which can be inflammation or infection. Unfortunately, the physician's clinical diagnosis of bacteremia lacks both sensitivity and specificity.35 Blood culture acquisition is a simple and basic diagnostic procedure routinely used in clinical practice that yields essential information for the evaluation of various infectious diseases.6 Positive blood cultures can demonstrate not only an infectious cause of disease but also the microbiological response to antibiotic therapy.7 However, studies have reported that 35% to 50% of positive blood cultures are falsely positive owing to contamination.711 False‐positive cultures may lead to the use of inappropriate or unnecessary antibiotics, additional testing and consultation, and prolonged hospitalizations that increase patient care costs.9, 12

Nursing staff caring for patients are generally able to assess oral intake, general clinical state, and care requirements. Moreover, the nursing staff are often able to identify problems with patients before physicians.13 In Japan, nurse‐assessed food consumption of every meal is standardized, and is frequently regarded as an indicator of the patient's clinical status, akin to a vital sign. In this context, we hypothesized that quantitative variations in food consumption could accurately distinguish those patients with or without bacteremia.

MATERIALS AND METHODS

RESULTS

During the study period, 851 patients aged 16 to 99 years (66.8 16.6), were eligible for inclusion (Figure 1). Baseline characteristics of the subjects are given in Table 1. The mean body temperature ( standard deviation [SD]) was 38.1 1.1C, and the mean CRP level was 8.7 8.1 mg/dL. The results show that the patients had at least 2 SIRS criteria with elevations in temperature and heart rate. Of the 851 patients entered into the study, only 122 (14.3%) had positive blood cultures. Of these, 75 patients (8.8%) were considered to have true‐positive blood culture. In this study, blood cultures were taken at the time of onset of fever, whether that was a new inpatient admission, or during the course of an admission to the hospital. On average, blood cultures were drawn 12 days after admission (SD, 5.6 days). Despite the variation in onset of fever, the inverse relationship of blood culture positivity to decreased food consumption held true (data not shown). Gram‐positive and Gram‐negative organisms were obtained in near equal amounts. The main pathogens recovered from the true‐positive blood cultures were Gram‐positive cocci (26 patients [34.7% in true‐positive blood cultures]), and Gram‐negative bacilli (46 patients [61.3%]), as shown in Table 1. The underlying clinical diagnosis included 28 urinary tract infections; 9 catheter‐associated infections; 5 cases each of pneumonia and abscess; 3 cases of phlebitis; 2 cases of meningitis and osteomyelitis; 1 case each of infective endocarditis, decubitus ulcer, and pelvic infection; and 17 cases of infection with an unknown focus.

Figure 1

Low, moderate, and high food consumption groups consisted of 344 patients (40.4%), 152 patients (17.9%), and 354 patients (41.7%), respectively (Table 1). Of these, 63 patients, 6 patients, and 6 patients had bacteremia in the low, moderate, and high food consumption group, respectively. In order to distinguish those patients who had decreased food consumption compared to almost normal food consumption, low and moderate food consumption groups were combined and compared to the high food consumption group. Comparison of the combined low and moderate food consumption group versus the high food consumption group revealed a sensitivity of 92.0% and a negative predictive value of 98.3% for excluding true bacteremia. Conversely, the specificity (45.1%) and the positive predictive value (13.9%) were poor.

In the univariate analysis, the following variables were not statistically significantly associated with true bacteremia: age, heart rate, and leukocyte counts. Significant univariate predictors of bacteremia and their associated cut‐off points were body temperature of 36 or 38C (odds ratio [OR], 2.5; 95% CI, 1.54.4), CRP 10.0 mg/dL (OR, 2.0; 95% CI, 1.23.2), and food consumption (OR, 8.5; 95% CI, 3.818.6) (Table 2). There was no evidence of colinearity. In the final stepwise logistic regression (Table 3), the significant predictors of bacteremia were body temperature of 36 or 38C (OR, 2.4; 95% CI, 1.44.2; P = 0.002), C‐reactive protein of 10.0 mg/dL (OR, 1.9; 95% CI, 1.23.0; P = 0.011), and food consumption (OR, 7.5; 95% CI, 3.416.6; P < 0.001). We identified only 6 patients with bacteremia in the high food consumption group. Three of the patients had been previously treated with antibiotics for conditions including infective endocarditis, osteomyelitis, and myelodysplasic syndrome.

On further analysis, we excluded patients who had received antibiotics before blood culture acquisition. There were 661 patients in this subanalysis. Low, moderate, and high food consumption groups consisted of 282 patients (41.4%), 118 patients (17.3%), and 261 patients (38.3%), respectively. Of these, 50 patients (17.7%), 5 patients (4.2%), and 4 patients (1.5%) had bacteremia in the low, moderate, and high food consumption groups, respectively. The sensitivity and negative predictive values were 93.2% and 98.5%, respectively. In the stepwise logistic regression, significant predictors of bacteremia were body temperature of 36 or 38C (OR, 3.0; 95% CI, 1.55.6; P = 0.001), CRP 10.0 mg/dL (OR, 2.1; 95% CI, 1.23.7; P = 0.006), and food consumption (OR, 9.3; 95% CI, 3.326.1; P < 0.001).

DISCUSSION

We found that in a group of 851 Japanese patients who were suspected with bacterial infection, the estimated food consumption was negatively associated, both significantly and independently, with the subsequent isolation of microorganisms from their blood cultures. If validated in other studies, this simple rule of thumb can provide the clinician with reasonable confidence that a febrile patient has a low probability of being bacteremic, as long as the appetite remains normal. Both the sensitivity and the negative predictive value were extremely high at 92.3% and 98.3%, respectively, suggesting that adequate oral intake is a strong marker against the presence of bacteremia. In this study, it was the strongest predictor of bacteremia in multivariate analysis. After including only antibiotic‐naive patients, the sensitivity and the negative predictive values were 93.2% and 98.5%, respectively. Administration of antibiotics may lead to improved appetite in febrile patients despite bacteremia in the presence of fever, and therefore, inquiring about recent or current antimicrobial usage should be a requirement when considering oral intake as an indicator of bacteremia.

Our study has limitations. Since we did not make treatment decisions based on oral intake, we cannot conclude that it is safe to withhold antibiotic treatment on the basis of food intake alone. Additionally, this study would need to be repeated across many different age groups and racial groups to ensure applicability to the general population. It is also unknown whether this rule would be applicable to patients with underlying immunosuppression. Finally, although inter‐rater reliability was high in our center, nurses in other settings may not be as diligent in their assessment of food consumption. The high inter‐assessor reliability in our setting, however, suggests that objective assessment of food intake can be performed reliably in settings in which accurate documentation of food consumption is expected.

In summary, we found that normal food intake was strongly and negatively associated with bacteremia in febrile patients. This observation, if validated in other settings, may serve as a simple aid to assist in the clinical diagnosis of bacteremia or for recruitment of patients with a high likelihood of bacteremia into clinical trials.

Fever is a nonspecific phenomenon that can result from many inciting causes such as infection, inflammation, malignancy, thromboembolic disease, drugs, and endocrine disease. In hospitalized patients, one of the most important clinical considerations is bacteremia. Although vital signs compose 3 of the 4 current criteria for the diagnosis of Systemic Inflammatory Response Syndrome (SIRS),1, 2 they contribute little to the diagnosis of the cause, which can be inflammation or infection. Unfortunately, the physician's clinical diagnosis of bacteremia lacks both sensitivity and specificity.35 Blood culture acquisition is a simple and basic diagnostic procedure routinely used in clinical practice that yields essential information for the evaluation of various infectious diseases.6 Positive blood cultures can demonstrate not only an infectious cause of disease but also the microbiological response to antibiotic therapy.7 However, studies have reported that 35% to 50% of positive blood cultures are falsely positive owing to contamination.711 False‐positive cultures may lead to the use of inappropriate or unnecessary antibiotics, additional testing and consultation, and prolonged hospitalizations that increase patient care costs.9, 12

Nursing staff caring for patients are generally able to assess oral intake, general clinical state, and care requirements. Moreover, the nursing staff are often able to identify problems with patients before physicians.13 In Japan, nurse‐assessed food consumption of every meal is standardized, and is frequently regarded as an indicator of the patient's clinical status, akin to a vital sign. In this context, we hypothesized that quantitative variations in food consumption could accurately distinguish those patients with or without bacteremia.

MATERIALS AND METHODS

RESULTS

During the study period, 851 patients aged 16 to 99 years (66.8 16.6), were eligible for inclusion (Figure 1). Baseline characteristics of the subjects are given in Table 1. The mean body temperature ( standard deviation [SD]) was 38.1 1.1C, and the mean CRP level was 8.7 8.1 mg/dL. The results show that the patients had at least 2 SIRS criteria with elevations in temperature and heart rate. Of the 851 patients entered into the study, only 122 (14.3%) had positive blood cultures. Of these, 75 patients (8.8%) were considered to have true‐positive blood culture. In this study, blood cultures were taken at the time of onset of fever, whether that was a new inpatient admission, or during the course of an admission to the hospital. On average, blood cultures were drawn 12 days after admission (SD, 5.6 days). Despite the variation in onset of fever, the inverse relationship of blood culture positivity to decreased food consumption held true (data not shown). Gram‐positive and Gram‐negative organisms were obtained in near equal amounts. The main pathogens recovered from the true‐positive blood cultures were Gram‐positive cocci (26 patients [34.7% in true‐positive blood cultures]), and Gram‐negative bacilli (46 patients [61.3%]), as shown in Table 1. The underlying clinical diagnosis included 28 urinary tract infections; 9 catheter‐associated infections; 5 cases each of pneumonia and abscess; 3 cases of phlebitis; 2 cases of meningitis and osteomyelitis; 1 case each of infective endocarditis, decubitus ulcer, and pelvic infection; and 17 cases of infection with an unknown focus.

Figure 1

Low, moderate, and high food consumption groups consisted of 344 patients (40.4%), 152 patients (17.9%), and 354 patients (41.7%), respectively (Table 1). Of these, 63 patients, 6 patients, and 6 patients had bacteremia in the low, moderate, and high food consumption group, respectively. In order to distinguish those patients who had decreased food consumption compared to almost normal food consumption, low and moderate food consumption groups were combined and compared to the high food consumption group. Comparison of the combined low and moderate food consumption group versus the high food consumption group revealed a sensitivity of 92.0% and a negative predictive value of 98.3% for excluding true bacteremia. Conversely, the specificity (45.1%) and the positive predictive value (13.9%) were poor.

In the univariate analysis, the following variables were not statistically significantly associated with true bacteremia: age, heart rate, and leukocyte counts. Significant univariate predictors of bacteremia and their associated cut‐off points were body temperature of 36 or 38C (odds ratio [OR], 2.5; 95% CI, 1.54.4), CRP 10.0 mg/dL (OR, 2.0; 95% CI, 1.23.2), and food consumption (OR, 8.5; 95% CI, 3.818.6) (Table 2). There was no evidence of colinearity. In the final stepwise logistic regression (Table 3), the significant predictors of bacteremia were body temperature of 36 or 38C (OR, 2.4; 95% CI, 1.44.2; P = 0.002), C‐reactive protein of 10.0 mg/dL (OR, 1.9; 95% CI, 1.23.0; P = 0.011), and food consumption (OR, 7.5; 95% CI, 3.416.6; P < 0.001). We identified only 6 patients with bacteremia in the high food consumption group. Three of the patients had been previously treated with antibiotics for conditions including infective endocarditis, osteomyelitis, and myelodysplasic syndrome.

On further analysis, we excluded patients who had received antibiotics before blood culture acquisition. There were 661 patients in this subanalysis. Low, moderate, and high food consumption groups consisted of 282 patients (41.4%), 118 patients (17.3%), and 261 patients (38.3%), respectively. Of these, 50 patients (17.7%), 5 patients (4.2%), and 4 patients (1.5%) had bacteremia in the low, moderate, and high food consumption groups, respectively. The sensitivity and negative predictive values were 93.2% and 98.5%, respectively. In the stepwise logistic regression, significant predictors of bacteremia were body temperature of 36 or 38C (OR, 3.0; 95% CI, 1.55.6; P = 0.001), CRP 10.0 mg/dL (OR, 2.1; 95% CI, 1.23.7; P = 0.006), and food consumption (OR, 9.3; 95% CI, 3.326.1; P < 0.001).

DISCUSSION

We found that in a group of 851 Japanese patients who were suspected with bacterial infection, the estimated food consumption was negatively associated, both significantly and independently, with the subsequent isolation of microorganisms from their blood cultures. If validated in other studies, this simple rule of thumb can provide the clinician with reasonable confidence that a febrile patient has a low probability of being bacteremic, as long as the appetite remains normal. Both the sensitivity and the negative predictive value were extremely high at 92.3% and 98.3%, respectively, suggesting that adequate oral intake is a strong marker against the presence of bacteremia. In this study, it was the strongest predictor of bacteremia in multivariate analysis. After including only antibiotic‐naive patients, the sensitivity and the negative predictive values were 93.2% and 98.5%, respectively. Administration of antibiotics may lead to improved appetite in febrile patients despite bacteremia in the presence of fever, and therefore, inquiring about recent or current antimicrobial usage should be a requirement when considering oral intake as an indicator of bacteremia.

Our study has limitations. Since we did not make treatment decisions based on oral intake, we cannot conclude that it is safe to withhold antibiotic treatment on the basis of food intake alone. Additionally, this study would need to be repeated across many different age groups and racial groups to ensure applicability to the general population. It is also unknown whether this rule would be applicable to patients with underlying immunosuppression. Finally, although inter‐rater reliability was high in our center, nurses in other settings may not be as diligent in their assessment of food consumption. The high inter‐assessor reliability in our setting, however, suggests that objective assessment of food intake can be performed reliably in settings in which accurate documentation of food consumption is expected.

In summary, we found that normal food intake was strongly and negatively associated with bacteremia in febrile patients. This observation, if validated in other settings, may serve as a simple aid to assist in the clinical diagnosis of bacteremia or for recruitment of patients with a high likelihood of bacteremia into clinical trials.