Impact of Fever on Outcome in Patients With Stroke and Neurologic Injury
A Comprehensive Meta-Analysis
Background and Purpose— Many studies associate fever with poor outcome in patients with neurological injury, but this relationship is blurred by divergence in populations and outcome measures. We sought to incorporate all recent scholarship addressing fever in brain-injured patients into a comprehensive meta-analysis to evaluate disparate clinical findings.
Methods— We conducted a Medline search for articles since January 1, 1995 (in English with abstracts, in humans) and hand searches of references in bibliographies and review articles. Search terms covered stroke, neurological injury, thermoregulation, fever, and cooling. A total of 1139 citations were identified; we retained 39 studies with 67 tested hypotheses contrasting outcomes of fever/higher body temperature and normothermia/lower body temperature in patients with neurological injury covering 14431 subjects. A separate meta-analysis was performed for each of 7 outcome measures. Significance was evaluated with Zc developed from probability values or t values. Correlational effect size, r(es), was calculated for each study and used to derive Cohen’s d unbiased combined effect size and relative risk.
Results— Fever or higher body temperature was significantly associated with worse outcome in every measure studied. Relative risk of worse outcome with fever was: mortality, 1.5; Glasgow Outcome Scale, 1.3; Barthel Index, 1.9; modified Rankin Scale, 2.2; Canadian Stroke Scale, 1.4; intensive care length of stay, 2.8; and hospital length of stay, 3.2.
Conclusions— In the pooled analyses covering 14431 patients with stroke and other brain injuries, fever is consistently associated with worse outcomes across multiple outcome measures.
Fever is a common condition in patients with stroke and other brain injuries. Hyperthermia appears to correlate with poor outcome in these patients, although a direct causative link has not been established. The impact of fever on patients in a neurocritical care unit has been evaluated; after controlling for severity of illness, diagnosis, age, and complications, fever was found to be strongly associated with an increased length of intensive care unit (ICU) and hospital stay as well as a higher mortality rate and worse overall outcome.1 These findings are consistent with the meta-analysis by Hajat and colleagues focused on stroke mortality and generalized morbidity.2 In this study, we expand Hajat’s analysis to assess the significance of the relationship between fever and outcome across the full range of neurological injury, including stroke, and its magnitude in distinct measures of clinical, functional, and economic outcome.
Excellent biological arguments exist for a direct impact of fever on neurological outcome after brain injury. On a local level, fever results in the following: (1) elevated levels of excitatory amino acids (eg, glutamate and dopamine), free radicals, lactic acid, and pyruvate3; (2) increased ischemic depolarizations; (3) blood-brain barrier breakdown; (4) impaired enzymatic function; and (5) reduced cytoskeletal stability. Globally, these events lead to both cerebral edema, potentially reducing cerebral perfusion pressure, and larger volumes of ischemic injury.4,5 The inciting etiology of the brain injury appears almost immaterial when considering the aforementioned effects, because ischemic stroke,6–8 subarachnoid hemorrhage,9,10 intracerebral hemorrhage,11,12 traumatic brain injury,13 and global ischemic injury from cardiac arrest14 have all been noted to be impacted by fever in these ways.
There is a large body of research investigating various aspects of the relationship between hyperthermia—compared with normothermia—and outcome in patients with ischemic, hemorrhagic, and traumatic brain injuries. The studies are predominantly observational and retrospective, test a multitude of hypotheses, and involve a large number of intervening variables, a host of different temperatures indicative of fever, different times of onset and durations of fever, and a multiplicity of measures used to ascertain outcome. Taken singly, they provide limited and sometimes uncertain guidance. The impact of fever is significant in many studies, although some studies have found no significance. Yet despite much variance, the literature is characterized by the consistent suggestion that fever, when it is significant, contributes to problems in patients with all types of neurological injury. A meta-analysis, although it does not substitute for a randomized, controlled trial, is an approach that is well suited in this circumstance.15 It provides a way to pull together these single studies often examining narrowly drawn aspects of this association in highly individual ways to systematically test larger questions: does fever contribute to worse outcomes across the full spectrum of brain-injured patients? Is the effect of fever significant and large enough to be clinically important in all commonly measured aspects of patient outcome?
Materials and Methods
A Medline search for articles published since January 1, 1995, for articles (in English with abstracts, in humans) pertaining to fever in neurological patients was completed in March 2006 using the following search terms: [(head OR brain OR cerebral OR neuro OR cranial OR subdural OR epidural OR arachnoid) AND (fever OR temperature OR hyperthermia OR pyrexia) AND (tumor OR trauma OR injury OR damage OR ischemia OR hemorrhage OR tbi OR lesion OR infarct OR coma OR edema OR hematoma OR contusion OR icp OR sah OR ich OR sdh OR edh) AND (induced hypothermia OR therapeutic hypothermia OR hypothermic therapy OR thermoregulation OR thermoregulatory OR temperature control OR heat exchange OR cooling OR prognosis OR prognostic OR predictor OR predict OR outcome OR incidence)]. We included terms pertaining to hypothermia to ensure identification of all studies containing data on relevant body temperature ranges, even if a study’s overt topic was not relevant to our analyses. The quality standard for inclusion was a complete clinical study published in a Medline-abstracted journal. A total of 1139 citations was identified. The reference lists in 35 systematic reviews identified among the original 1139 citations were hand-searched for additional studies of relevance to these analyses as were the bibliographies of all selected studies. Case studies, mechanistic studies of cooling technologies, reviews, comments, and studies on hypothermia in patients with cardiac arrest or lacking data on fever were eliminated. In all, 45 studies, selected without regard to the specific outcome measure(s) used, were identified as potentially appropriate (citations in Supplemental Table I⇓, available online at http://stroke.ahajournals.org). Three studies involving cooling technologies were retained because they provided comparative data on afebrile patients and patients whose body temperatures were consistent with definitions of fever used in other studies (>37.2°C, Mayer et al; ≥38C°, Qui et al; >38°C, Yamamoto et al). Our minimum requirement for pooling was a least 3 studies using the same outcome measure. Five studies were eliminated because the minimum requirement for pooling was not met, and one study was withdrawn because it lacked sufficient data to relate a single source of body temperature to outcome. Thirty-nine studies (87%) were ultimately included in our study as detailed in the flow sheet (Figure 1). Twenty-two studies (56%) were prospective. Two studies met criteria for evidence Grade 1b (individual randomized, controlled trial with a narrow confidence interval, 35 studies were Grade 2b (individual cohort study/low-quality randomized, controlled trial), and 2 studies were graded 4 (case-series/poor-quality cohort or case-control), as identified in Table 1.16 Final selection of articles was made independently by 3 of the investigators. Data were extracted by a single investigator and reviewed independently by 2 others. All disagreements were resolved by consensus.
A separate meta-analysis was performed for each outcome measure in which our 3-study minimum was met. Only one meta-analysis contained less than 5 studies. Where a study had 2 hypotheses relevant to a given outcome measure, both were evaluated. In the mortality meta-analysis, in which some studies had 3 or 4 hypotheses assessing small variations, the single hypothesis with the most distant mortality point and the least restrictive hyperthermia timing was evaluated. In all, we evaluated 67 hypotheses from 39 studies addressing the following widely used measures of clinical, functional, and economic outcome: mortality, Glasgow Outcome Scale (GOS), Barthel Index (BI), modified Rankin Scale (mRS), Canadian Stroke Scale (CSS), ICU length of stay (LOS), and hospital LOS. Sample sizes ranged from 38 to 4295, as shown in Table 1, with a total of 14431 patients. The clinical populations of these studies overlapped. To minimize selection bias, we included studies that combined ischemic and hemorrhagic stroke and studies of neurological ICU populations that also included patients with traumatic brain injury (TBI) as well as studies examining fever in targeted stroke and TBI populations.
Our intention to include as many studies as possible, despite great variation in individual study statistics, necessitated a flexible statistical approach. For each outcome measure, a combined test (Zc) was developed from probability values (Stouffer’s approach) or t values (Winer’s approach; whichever was available in the source study) and was used to determine whether there was a significant difference between the fever and nonfever groups. We calculated individual effect size r(es) for each study using Rosenthal’s correlational approach17 and used it to derive unbiased Cohen’s d from which combined effect size was calculated for each pooled analysis. Combining effect sizes is preferable to combining probabilities from separate studies because it unambiguously adjusts for different sample sizes in the combined analysis.18
In addition, we derived more clinically relevant relative risk (RR) statistics for each study and for each meta-analysis. For each study, available statistics such as odds ratios, probability values, and t values were used to calculate the proportion of each group (febrile/afebrile) having a good or bad outcome, and the results were presented in a Binomial Effect-Size Display table from which relative risk for each meta-analysis was derived. Rosenthal and Rubin’s Binomial Effect-Size Display approach is consistent with the dichotomized outcome reporting used in many of the studies, and it avoided excluding studies solely on the basis of statistical presentation. CIs for RR are not shown because the Binomial Effect-Size Display method for calculating RR, necessitated by the available statistics, produces a CI involving a fixed lower bound that is not directly comparable to CI around an RR calculated with more common approaches. (See Supplemental Table II⇓⇓⇓⇓⇓, available online at http://stroke.ahajournals.org, for counter effect size provided as an alternative approach to CI around correlational effect size.) Finally, homogeneity of effect size was assessed with a χ2 test using Cohen’s d.
Description of Data Variability
Multivariate studies included a wide array of moderating or mediating variables in addition to body temperature. Studied factors included patient demographics, comorbidities, severity indicators, clinical indicators, and timing of fever onset. Covariates and results for each tested hypothesis are shown in Supplemental Table II⇑⇑⇑⇑⇑.
As noted previously, the included articles used a range of statistical methods depending on the variables studied and on how the body temperature variable was structured. Most studies examined body temperature as a dichotomized (eg, febrile versus afebrile) or categorical (level of fever) variable, whereas some treated body temperature as a continuous variable. Thus, we refer to “fever/higher body temperature” in reporting findings across these approaches. Of studies defining fever, 13 studies used 37.5°C as the cutoff point; values from other studies ranged from 37.0 (tympanic) to 39.0°C (core), listed in Supplemental Table II⇑⇑⇑⇑⇑.
Studies that provided sufficient detail for GOS or mRS to permit separate assessment of mortality were included in the mortality meta-analysis even if mortality was not overtly discussed in the original study. Whenever a study in the meta-analysis included a pediatric population, the meta-analysis was conducted with and without the pediatric study. In no case did inclusion of a pediatric study make a statistically significant difference in the results.
Collectively, the meta-analyses presented a consistent result. In each of the 7 outcome measures evaluated, the meta-analysis indicated that fever/higher body temperature was significantly associated with worse outcome as indicated by higher mortality rates, greater disability, more dependence, worse functional outcome, greater severity, and longer stays in the hospital and ICU. The size of the effect of fever/higher body temperature on outcome in each meta-analysis ranged from moderately small (0.26 for GOS) to large (over 0.8 for mRS, ICU LOS, and hospital LOS) per Cohen’s interpretation,19 as shown in Table 2. The RR calculations for each meta-analysis (Table 2) showed results consistent with the effect size analysis. The probability of a poorer outcome among patients with hemorrhagic or ischemic stroke or traumatic brain injury who had fever/higher body temperatures ranged from 1.3 for GOS to 3.2 times greater for hospital LOS than the probability among such patients without fever or with lower body temperatures.
Figure 2 shows the meta-analyses RR results and the individual RR result for each component study grouped by outcome measure and the clinical subgroup that best categorizes the study’s population. Studies involving ischemic stroke exclusively were more consistent in their RR results than were studies also including or focusing exclusively on hemorrhagic stroke or traumatic brain injury. Mortality and GOS show the tightest clustering of individual RR around the meta-analysis results, illustrating more consistency of the relationship between fever/higher body temperature and outcomes across clinical groupings for these 2 measures compared with some of the other outcome measures. However, the overall results show that the association of fever with poorer clinical outcomes cuts across all types of neurological injury and is observed in every outcome measure analyzed.
Heterogeneity of effect size, measured using Cohen’s d, was insignificant in 3 meta-analyses (mortality, GOS, and CSS) and present in 4 (ICU LOS, hospital LOS, mRS, and BI). Statistical heterogeneity is reasonably common in meta-analyses using studies with divergent structures but always warrants further examination. In these meta-analyses, there are a variety of possible contributors to heterogeneity. In the LOS analyses (ICU and hospital LOS), some studies were testing the impact of fever on LOS, whereas others used fever as the dependent variable.20,21 Furthermore, a continuous variable such as LOS can produce larger effect sizes and therefore more potential variance than dichotomous variables or dichotomized ordinal scales. For studies examining the mRS, differences in the timing and duration of the body temperature reading(s) being assessed may have contributed to heterogeneity of effect size. In the BI meta-analysis, several studies investigated essentially the same hypothesis with different statistical methods (nonparametric tests of ordinal [Mann-Whitney U or Kruskal-Wallis] or nominal [Fisher exact test] data and/or logistic regression analysis). In addition, there was variation in cut scores for good and poor outcomes across studies as well as variation in timing of outcome measurement.
Despite heterogeneity, the relationship between fever and negative outcome was statistically significant and consistent in direction in all 7 meta-analyses.
For each measure of clinical outcome, the studies singly and collectively demonstrate a clear link between fever and worse outcome. Among 67 hypotheses in 39 studies evaluated in these meta-analyses, the relationship between fever/higher body temperature and worse outcome was confirmed as significant in 53 hypotheses, found to be statistically insignificant in 13, and significantly rejected in one, as detailed in Supplemental Table II⇑⇑⇑⇑⇑. Independent of the origin of the neurological injury (ischemic, hemorrhagic, or traumatic), fever was associated with negative outcomes. The meta-analyses confirm that the association between fever and negative outcomes extends across all dimensions of outcome for which studies were available to analyze. Although the size of the effect varies, the range of RR indicates that fever in patients with neurological injury is a condition of clinical importance.
Relative Risk by Outcome Measure
As shown in Table 2, the largest values for RR in the meta-analyses were found in the 2 LOS analyses. Of note, several LOS studies measured fever over the entire duration of the ICU stay so that patients with a longer stay had a greater chance of fever being identified. However, several studies measuring LOS found a significant association between outcome in patients with stroke and body temperature measured on admission, suggesting that the relationship between longer LOS and fever cannot be attributed solely to treatment effect.6,7,22–24 Although the magnitude of relative risk in ICU LOS and hospital LOS may be affected by studies with variable periods of measurement, and by the fact that LOS is a continuous variable, the direction of these findings is consistent with the other meta-analyses.
The smallest values for RR come from the GOS (RR=1.3) and CSS (RR=1.4). These results are significant but not large. However, they do suggest that fever is associated with greater neurological dysfunction with neurological injury. The higher RR levels in conjunction with mortality and functional outcome measures provide confirmation that fever in neurological injury is a clinically important condition that needs to be definitively studied.
The question of whether fever actually causes worse outcomes or whether it is largely an effect of other causative factors, known or unknown, cannot be fully addressed in this type of analysis. However, the meta-analyses and component studies do contain useful insights on this point. Across the 39 studies we analyzed, including studies involving widely defined multivariate analyses (see Supplemental Table II⇑⇑⇑⇑⇑), findings on most covariates are consistently insignificant or mixed. No other covariate comes close to reaching the consistency of significance found with fever/higher body temperature.
Effect Size, Timing, and Clinical Condition
Individual researchers’ study designs as well as their findings reflect differences in the etiology of fever in varying types of neurological injury as well as the timing of fever measurements. In checking for possible sources of bias, we examined the distribution of the effect size of fever (irrespective of outcome measure) by the timing and/or duration of temperature measurement within these clinical groups. We found no observable trends large enough to warrant examination. We did, however, notice differences between clinical groups with regard to the timing of temperature measurement. Studies in our meta-analysis that evaluated broadly defined neurological populations with a mixture of TBI and stroke all elected to study fever throughout an ICU stay or extended time period. For the narrower clinical groups, 67% of the hypotheses in the hemorrhagic and all stroke studies evaluated results of temperatures taken on admission or within approximately the first 24 hours. For studies of ischemic stroke, 40% of the hypotheses focused on these early temperature measurements. In the TBI studies, only 20% of the studies focused on temperatures in the first 24 hours and none considered admission temperature alone. Differences in study design suggest that stroke researchers are more apt to focus on early fevers, whereas TBI researchers are interested in fever less immediately and for longer periods of time. These trends reflect the differing nature of the neurological injury; a majority of ischemic injury occurs within the first several hours, whereas in subarachnoid hemorrhage, for example, the period of brain injury that might be modified by body temperature may extend for several days. We believe that questions about when and for how long fever should be measured in various types of neurological injury is a fruitful area for future research.
Our study points to the timeliness and compelling justification for a major prospective study in neurologically injured patients to determine whether outcomes improve when fever is prevented or controlled. Findings of this meta-analysis suggest that such a study should include functional and economic outcome measures in addition to clinical ones and should be designed to yield guidance for practicing clinicians on the important questions of not only whether, but when and for how long, to maintain thermoregulation. It also remains to be seen what impact other variables of the fever may have, including fever severity, timing, and duration.
The following factors are limitations of this study: (1) there is a possible selection bias from the choice of published studies, in English; (2) 17 studies were not from prospective trials; (3) 4 of the 5 studies using the CSS were done by the same group of researchers; (4) studies used different definitions of fever and different methods of measuring temperature, introducing a potential measurement bias; (5) studies dichotomized the same outcome measure differently, but the extent to which these differences fully compensated for differences in population severity cannot be determined; (6) an insufficient number of randomized, controlled trials was available to permit pooling by study type; (7) not all studies provided exact probability values; (8) probability values or coefficients were not always reported on insignificant results; (9) not all articles contained sufficient statistics for Stouffer’s Zc, necessitating the use of a different combined measure of statistical significance for some analyses; and (10) heterogeneity of effect size was found in 4 of the 7 meta-analyses, and it was not possible to test for all root causes of heterogeneity.
In the most comprehensive analysis undertaken to date, incorporating 39 studies covering 14431 patients with stroke and other brain injuries, fever/higher body temperature is consistently associated with worse outcome regardless of the outcome measure used. Fever was consistently associated with damaging effects, whether the neurological injury was ischemic, hemorrhagic, or traumatic in origin and whether the outcome being measured was clinical, functional, or economic. A major prospective study is clearly warranted to confirm whether aggressive efforts to prevent and control fever in neurologically injured patients will yield improvements in these outcomes.
S.E.F., N.L.R., M.O., and G.C.U. report financial support from Medivance, Inc under an agreement that the study be conducted independently to reduce funding bias. Accordingly, the study was designed, conducted, analyzed, interpreted, and written by investigators independent of Medivance and was not sent to agents of Medivance for prepublication review or approval.
- Received March 27, 2008.
- Accepted April 11, 2008.
Hajat C, Hajat S, Sharma P. Effects of poststroke pyrexia on stroke outcome: a meta-analysis of studies in patients. Stroke. 2000; 31: 410–414.
Stocchetti N, Protti A, Lattuada M, Magnoni S, Longhi L, Ghisoni L, Egidi M, Zanier ER. Impact of pyrexia on neurochemistry and cerebral oxygenation after acute brain injury. J Neurol Neursurg Psychiatry. 2005; 76: 1135–1139.
Baena RC, Busto R, Dietrich WD, Globus MY, Ginsberg MD. Hyperthermia delayed by 24 hours aggravates neuronal damage in rat hippocampus following global ischemia. Neurology. 1997; 48: 768–773.
Kammersgaard LP, Jørgensen HS, Rungby JA, Reith J, Nakayama H, Weber UJ, Houth J, Olsen TS. Admission body temperature predicts long-term mortality after acute stroke: the Copenhagen Stroke Study. Stroke. 2002; 33: 1759–1762.
Boysen G, Christensen H. Stroke severity determines body temperature in acute stroke. Stroke. 2001; 32: 413–417.
Oliveira-Filho J, Ezzeddine MA, Segal AZ, Buananno FS, Chang Y, Ogilvy CS, Rordorf G, Schwamm LH, Koroshetz WJ, McDonald CT. Fever in subarachnoid hemorrhage. Relationship to vasospasm and outcome. Neurology. 2001; 56: 1299–1304.
Schwarz S, Häfner K, Aschoff A, Schwab S. Incidence and prognostic significance of fever following intracerebral hemorrhage. Neurology. 2000; 54: 354–361.
Leira R, Dávalos A, Silva Y, Gil-Peralta A, Tejada J, Garcia M, Castillo J. Early neurologic deterioration in intracerebral hemorrhage: predictors and associated factors. Neurology. 2004; 63: 461–467.
Wolf FM. Meta-Analysis: Quantitative Methods for Research Synthesis. Newbury Park, CA: Sage Publications; 1986.
Phillips B, Ball C, Sackett D, Badenoch D, Straus S, Haynes B, Dawes M. Oxford Centre for Evidence-Based Medicine Levels of Evidence. 2001. Available at: www.cebm.net/levels_of_evidence.asp#levels. Accessed September 27, 2005.
Rosenthal R, Rosnow RL, Rubin DB. Contrasts and Effect Sizes in Behavioral Research: A Correlational Approach. Cambridge, UK: Cambridge University Press; 2000.
Becker BJ. Combining significance levels. In: Cooper HM, Hedges LV, eds. The Handbook of Research Synthesis. New York: Russell Sage; 1994.
Cohen J. Statistical Power Analysis for Behavioral Sciences,rev ed. New York: Academic Press; 1977.
Commichau C, Scarmeas N, Mayer SA. Risk factors for fever in the neurologic intensive care unit. Neurology. 2003; 60: 837–841.
Suzuki S, Kelley RE, Dandapani BK, Reyes-Iglesias Y, Dietrich WD, Duncan RC. Acute leukocyte and temperature response in hypertensive intracerebral hemorrhage. Stroke. 1995; 26: 1020–1023.
Wang Y, Lim LL, Levi C, Heller R, Fisher J. Influence of admission body temperature on stroke mortality. Stroke. 2000; 31: 404–409.