Chest Email Content Delivery
HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
QUICK SEARCH:   [advanced]


     

Guest Access | Sign In via User Name/Password
This Article
Right arrow Abstract Freely available
Right arrow Full Text (PDF) Free
Right arrow Submit a response
Right arrow Alert me when this article is cited
Right arrow Alert me when eLetters are posted
Right arrow Alert me if a correction is posted
Right arrow Citation Map
Services
Right arrow Email this article to a friend
Right arrow Similar articles in this journal
Right arrow Similar articles in ISI Web of Science
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Add to My Personal Article Archive
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via ISI Web of Science (18)
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Pascual, F. E.
Right arrow Articles by Wachter, R. M.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Pascual, F. E.
Right arrow Articles by Wachter, R. M.
(Chest. 2000;117:503-512.)
© 2000 American College of Chest Physicians

Assessment of Prognosis in Patients With Community-Acquired Pneumonia Who Require Mechanical Ventilation*

Fred E. Pascual, MD; Michael A. Matthay, MD, FCCP; Peter Bacchetti, PhD and Robert M. Wachter, MD

* From the Cardiovascular Research Institute (Dr. Matthay), the Department of Epidemiology and Biostatistics (Dr. Bacchetti), and the Department of Medicine (Drs. Pascual and Wachter), University of California, San Francisco, San Francisco CA; and Pulmonary & Critical Care Medicine (Dr. Pascual), Harborview Medical Center, Seattle, WA. Supported by the UCSF Department of Medicine and NIH grant HL51856.

Correspondence to: Fred E. Pascual, MD, Harborview Medical Center, Pulmonary & Critical Care Medicine, 325 9th Ave, Box 359762, Seattle, WA 98104; e-mail: feppccm{at}usa.net


   Abstract
TOP
 Abstract
Introduction
 Materials and Methods
 Results
 Discussion
 References
 
Study objectives: Knowing that mortality is high in patients who require mechanical ventilation patients with community-acquired pneumonia (CAP), we hypothesized that the severity of acute lung injury could be used along with nonpulmonary factors to identify patients with the highest risk of death. We formulated a prediction model to quantitate the risk of hospital mortality in this population of patients.

Design: Historical prospective study using data collected over the first 24 h of mechanical ventilation. We utilized a hypoxemia index—(1 - lowest [PaO2/PAO2]) x (minimum fraction of inspired oxygen to maintain PaO2 at > 60 mm Hg) x 100], where PAO2 is the alveolar partial pressure of oxygen—to grade the severity of acute lung injury on a scale from 0 to 100.

Setting: Tertiary care university hospital ICU.

Patients: One hundred forty-four adult patients mechanically ventilated for respiratory failure caused by CAP.

Measurements and results: Hospital mortality was 46% (n = 66). Multivariate logistic regression analysis revealed five independent predictors of hospital mortality: (1) the extent of lung injury assessed by the hypoxemia index; (2) the number of nonpulmonary organs that failed; (3) immunosuppression; (4) age > 80 years; and (5) medical comorbidity with a prognosis for survival < 5 years. At a 50% mortality threshold, the prediction model correctly classified outcome in 88% of cases. All patients with > 95% predicted probability of death died in hospital.

Conclusions: Based on clinical parameters measured over the first 24 h of mechanical ventilation, this model accurately identified critically ill, mechanically ventilated patients with CAP for whom prolonged intensive care may not be of benefit.

Key Words: hypoxemia index • intensive care • mechanical ventilation • outcome • pneumonia • prognosis


   Introduction
TOP
 Abstract
 Introduction
Materials and Methods
 Results
 Discussion
 References
 
Clinicians in the ICU must often distinguish between patients who have a reversible medical illness and those whose condition is so advanced that aggressive medical care will be of little impact. Such knowledge can clearly influence overall treatment strategies as well as aid in discussions about prognosis.

Community-acquired pneumonia (CAP) represents one of the most common causes of ICU admission.1 Prior investigations of CAP in the ICU have shown that the requirement for mechanical ventilation is associated with increased mortality compared with nonventilated patients.2 3 4 5 6 7 8 However, among the subgroup of patients requiring mechanical ventilation, the severity of lung injury has not been evaluated in detail as a prognostic factor. Therefore, it is uncertain whether prognostic criteria in the literature (derived from mixed populations of ventilated and nonventilated patients) remain pertinent in the patient requiring mechanical ventilation. We hypothesized that the extent of lung injury is a cardinal determinant of outcome in patients with respiratory failure from community-acquired pneumonia, and formulated a hypoxemia index: (1 - lowest [PaO2/PAO2]) x (minimum FIO2 to maintain PaO2 at > 60 mm Hg) x 100, where FIO2 is the fraction of inspired oxygen and PAO2 is the alveolar partial pressure of oxygen, to measure the extent of lung injury. The advantage of this index over single measurements in time, such as the PaO2, alveolar-arterial oxygen pressure difference (P(A-a)O2), PaO2/FIO2, PaO2/PAO2, FIO2, and level of positive end-expiratory pressure (PEEP), is that it accounts for improvement or worsening of lung injury because it summarizes the degree of lung injury over a given time period.

We present a prediction rule to quantitate the risk of hospital mortality in mechanically ventilated patients with CAP. The rule utilizes information obtained over the first 24 h of mechanical ventilation. Early, accurate prognostic estimates may be useful for establishing treatment objectives if clinicians can identify patients with a very high mortality risk for whom prolonged ICU care may not be beneficial. Early knowledge of which patients are highly unlikely to survive may help alleviate physician, patient, and family anxiety arising from prognostic uncertainty in the first few days of critical illness, as well as help facilitate discussions about care near the end of life.

This study utilizes multivariate analysis, which adjusts for confounding variables while selecting for predictors that independently contribute to the discriminate power of the model. Only a few studies of severe CAP in the ICU have utilized multivariate analysis.2 3 7 Logistic regression is a form of multivariate analysis that yields a calculated probability of an outcome (in this study, hospital mortality) once the values of each independent variable are known. We hypothesized that these techniques could identify a subset of patients with severe CAP whose lung injury and other prognostic variables predicted a very low chance of survival. Furthermore, using receiver-operator curve receiver operating characteristics (ROC) analysis, we compared the performance of the prediction rule with other recognized ICU prediction models (simplified acute physiology score [SAPS] and acute physiology and chronic health evaluation [APACHE] II).


   Materials and Methods
TOP
 Abstract
 Introduction
 Materials and Methods
Results
 Discussion
 References
 
Patient Selection
We reviewed chart data from 144 consecutive admissions of adult patients with severe CAP requiring mechanical ventilation. The study was approved by our institution’s committee on human research. The patients were admitted to the ICU of Moffitt-Long Hospital between July 1994 and July 1997. Nine patients were excluded because of misplaced records. The main outcome measure was hospital mortality. Thirty-five predictive variables and antibiotic use, assessed over the first 24 h of mechanical ventilation, were compared in the surviving vs nonsurviving groups (Tables 1 , 2) . Selection of these variables was based on clinical experience as well as review of prior published data.2 3 7


View this table:
[in this window]
[in a new window]

  		Table 1.. Univariate Analysis, Significant Predictors*

 

View this table:
[in this window]
[in a new window]

  		Table 2.. Univariate Analysis, Nonsignificant Predictors

 
Definitions
CAP was defined as an acute respiratory illness acquired in a nonhospital setting and manifested by characteristic symptoms (cough, sputum production, dyspnea, pleuritic chest pain, or altered mental status) with the appearance of a new compatible radiographic infiltrate, accompanied by fever (> 38.5°C) or hypothermia (< 36.0°C), and/or an abnormally increased or decreased WBC count (normal range, 3,400 to 10,000 cells/mL). Aspiration as an overt cause was defined similarly, with the addition of a predisposing factor such as altered mental status, coma, drug or alcohol intoxication, seizure, vomiting prior to the onset of respiratory symptoms, the presence of an impaired gag reflex, or a witnessed aspiration event. All patients required mechanical ventilation for one or more of the following reasons: hypoxemic respiratory failure (generally, PaO2 < 60 mm Hg on maximal supplemental oxygen); hypercarbic respiratory failure (generally, pH < 7.25 with PCO2 > 50 mm Hg); or the inability to protect the airway due to depressed mental status. Immunosuppression was defined as the use of corticosteroids (prednisone >= 20 mg/d) or other immunosuppressive medications (cyclophosphamide, cyclosporine, etc.), HIV infection with CD4 count < 200, neutropenia with absolute neutrophil count < 1,000, or other medical disorders known to cause immune dysfunction, such as chronic granulomatous disease, chronic lymphocytic leukemia, or multiple myeloma. Medical comorbidities were defined as diabetes mellitus, alcoholism, IV drug use, COPD, asthma, chronic interstitial lung disease, chronic liver disease, coronary artery disease, congestive heart failure, systemic lupus erythematosus, chronic renal insufficiency or failure, cystic fibrosis, cerebrovascular disease, solid tumor, or hematologic malignancy. Examples of medical conditions with a median prognosis for survival < 5 years included cancer of the esophagus, stomach, biliary tract, pancreas, lung; metastatic cancer from any site; AIDS with a CD4 count < 2009 ; cirrhosis with ascites or varices10 ; oxygen-dependent COPD11 ; New York Heart Association Class III-IV congestive heart failure12 ; dialysis-dependent chronic renal failure13 ; multiple myeloma14 ; myelodysplastic syndrome15 ; and acute myelogenous leukemia.16

We recorded the lung injury score as defined by Murray et al,17 the SAPS according to Le Gall et al,18 and APACHE II according to Knaus et al.19 ARDS was defined as the presence of bilateral radiographic infiltrates accompanied by a PaO2/FIO2 ratio < 200.20 Definitions for nonpulmonary organ failure were as follows: (1) cardiovascular: mean arterial pressure < 60 mm Hg or need for vasopressors > 6 h; (2) acute renal failure: urine output <= 479 mL/24 h or <= 159 mL/8 h, serum BUN >= 100 mg/dL, or serum creatinine >= 3.5 mg/dL; (3) hematologic: WBC count < 1,000 cells/mL, platelets <= 20,000 cells/mL, or hematocrit <= 20; and (4) neurologic: Glasgow Coma Score < 6 in the absence of sedation. These criteria are identical to those of Knaus et al21 except for the added definition of pressor use > 6 h as a surrogate for cardiovascular failure. An etiologic organism was isolated if cultures from blood, BAL fluid, or pleural fluid grew a bacterial organism compatible with CAP, if sputum was positive on smear or culture for acid-fast bacilli or Nocardia, if organisms such as Aspergillus, Mucormycosis, Nocardia, Pneumocystis carinii, or acid-fast bacilli were isolated from BAL fluid, or if viral antigen tests (influenza, respiratory syncytial virus, adenovirus, or parainfluenza) of respiratory secretions were positive. Legionella antimicrobial coverage was defined as the use of erythromycin, doxycycline, or ciprofloxacin; pseudomonal coverage was defined as an appropriate antipseudomonal antibiotic initiated during the first 24 h of hospital admission or intubation. Bacteremia was defined as positive blood cultures drawn within the first 24 h of hospital admission.

Formulation of hypoxemia index
It was our clinical observation that the degree of lung injury varies markedly over the first 24 h of mechanical ventilation in CAP, when pneumonia is in evolution and antibiotics are taking effect. Therefore, we perceived a need to utilize a measure other than standard correlates of lung injury (PaO2/FIO2, PaO2/PAO2, FIO2, PEEP, presence of ARDS, and P(A-a)O2) which rely on a single measurement in time. We therefore formulated the hypoxemia index (see formula in second introductory paragraph), which integrates the most abnormal value of one parameter, the PaO2/PAO2, and the least abnormal value of another parameter, the FIO2, over a 24-h interval. In effect, the hypoxemia index summarizes the state of lung injury over a given time period, since it does not rely on single data points. The scale of measurement, 0 to 100 in proportion to severity, is easily interpreted.

Data Collection
The following data were recorded from ICU flowsheets within the first 24 h after initiation of mechanical ventilation: the lowest PaO2/PAO2 ratio, with alveolar PO2 defined as 713 (FIO2) - (PaCO2/0.8); the lowest PaO2/FIO2 ratio; the highest P(A-a)O2 (mm Hg), defined as 713 (FIO2) - (PaCO2/0.8) - PaO2; the lowest FIO2 to maintain PaO2 >= 60 mm Hg; the highest level of PEEP (cm H2O); the highest minute ventilation (E, in L/min); and the highest peak inspiratory pressure (cm H2O). Only the most abnormal values were recorded. Radiographic data were obtained from original reports dictated by a staff radiologist. Laboratory data, vital signs, Glasgow Coma Scores, urine outputs for calculation of SAPS, APACHE II scores, and the presence of nonpulmonary organ system failure were obtained from computer records and ICU flowsheets.

The derivation of the prediction rule by logistic regression was based on the premise that aggressive, ongoing intensive care was offered until death was imminent. Thus, 12 patients were excluded from this analysis. These patients included those who were extubated and transferred out of the ICU in stable condition and were given a "do not intubate, do not resuscitate" status for various reasons (often such decisions are made on the basis of subjective criteria, ie, quality of life judgments, and can be difficult to measure in a quantitative prediction rule), then died of recurrent respiratory failure when not reintubated (n = 11). In such cases, it was often not clear if survival would have been achieved by further trials of intensive care or mechanical ventilation. In addition, we excluded one patient who was initially intubated but from whom intensive care was subsequently withdrawn when the patient’s prior preferences against life-sustaining treatment became evident. Such exclusions are appropriate because in a clinical setting, early prognostic estimates should be based on the assumption that aggressive, intensive care will not be restricted.

Statistical Analysis
The software programs Statistica 5.1 (StatSoft; Tulsa, OK), SPSS 6.1.4 (SPSS Inc; Chicago, IL), SimStat for Windows (Provalis Research; Montreal, QC), and Stata 5.0 (Stat Corp; College Station, TX) were utilized for statistical analysis. Categorical variables were analyzed by 2 x 2 contingency tables using Fisher’s Exact Test. Continuous variables were compared by the Mann-Whitney U test. Significant univariate predictors (two-tailed p < 0.05) were entered into a forward stepwise logistic regression analysis, with criteria for entry and exit at p < 0.05 and < 0.10, respectively. Probabilities for death were correlated with a score derived by adding the products of each independent predictor with its regression coefficient, ie, Score = (b1x1) + (b2x2) + (b3x3) +... , and so on. Using the general logistic regression equation, the probability of death is related to the score by the following:

(1)
where b0 is the intercept; b1, b2, b3, etc, are regression coefficients; and x1, x2, x3, etc, are values for independent predictors. The model’s goodness of fit was assessed by the method of Hosmer and Lemeshow.22 To identify patients with an extremely poor chance of survival, we report the score above which the expected probability of death is > 95%. Cross-validation was used to assess variable selection and classification rates.23 ROC curves were constructed to compare sensitivity and specificity at all thresholds for our prediction model, our cross-validated model, SAPS, and APACHE II.

To further assess performance of the various measures of lung injury (hypoxemia index, PaO2/PAO2, PaO2/FIO2, lung injury score, presence of ARDS, P(A-a)O2, FIO2, and PEEP), we derived separate prediction models, starting from the 13 significant nonpulmonary univariate predictors (see Table 1 , A) and forcing one of the eight measures of lung injury into the model. Performance parameters (correct classification rates, goodness-of-fit tests, ROC curves, and identification of cases with > 95% mortality) were then compared.


   Results
TOP
 Abstract
 Introduction
 Materials and Methods
 Results
Discussion
 References
 
Subject Characteristics
The study population consisted of 144 patients; overall hospital mortality was 46% (n = 66). All data were obtained during the first 24 h of mechanical ventilation, except where noted. Tables 1 and 2 show baseline characteristics of the study population. An etiologic organism was identified in 30% of patients, with Streptococcus pneumoniae, Haemophilus influenzae, and Staphylococcus aureus being the most common isolates (Table 3) . Table 4 details antibiotic use at the start of mechanical ventilation.


View this table:
[in this window]
[in a new window]

  		Table 3.. Etiologic Organisms Isolated*

 

View this table:
[in this window]
[in a new window]

  		Table 4.. Antibiotic Use on Initiation of Mechanical Ventilation*

 
For nonsurvivors assigned > 95% probability of death (by our prediction rule, using the hypoxemia index as the measure of lung injury—see equation 2 , below), mean survival after intubation was 6 days (95% confidence interval [CI], 4 to 8 days). Table 5 lists modes of death in this population. The majority of nonsurvivors died while in the ICU (82%) or while intubated (74%). Intensive care (pressors and/or ventilatory support) was actively withdrawn prior to death in 78% of nonsurvivors. Among those who died while intubated, 92% had care actively withdrawn. No patient survived after a cardiopulmonary resuscitation attempt (n = 9). Thirteen percent died with significant CNS injury.


View this table:
[in this window]
[in a new window]

  		Table 5.. Modes of Death in 66 Patients Who Died in Hospital

 
Univariate Analysis
The 35 variables shown in Tables 1 and 2 , plus the 17 antibiotics shown in Table 4 , were examined for association with hospital nonsurvival. By univariate analysis, 21 of the variables in Tables 1 and 2 and two of the antibiotics listed in Table 4 were significantly associated with hospital mortality, at p < 0.05. These included all measures of lung injury (hypoxemia index, the minimum FIO2 to maintain PaO2 > 60 mm Hg, PEEP, the lung injury score, presence of ARDS, P(A-a)O2, PaO2/PAO2, and PaO2/FIO2 ratio) as well as immunosuppression, the number of nonpulmonary organs failed, isolation of an etiologic organism, male sex, age > 80 years, the presence of a preexisting medical condition with a median prognosis for survival < 5 years, pressor requirement, E, number of days in hospital prior to mechanical ventilation, bilateral radiographic involvement, the number of radiographic lobes involved, radiographic spread of infiltrates during the first 24 h of mechanical ventilation, and use of vancomycin or antifungal therapy. Aspiration was associated with decreased hospital mortality. Variables that were not associated with hospital mortality included peak inspiratory pressure, the number of preexisting comorbidities, prior residence in a nursing home, prior antibiotic treatment, bacteremia, pH, WBC count, serum bicarbonate level, age > 65 years, use of a pulmonary arterial catheter, and Legionella or Pseudomonas antibiotic coverage on admission to the hospital or at the time of intubation.

Multivariate Logistic Regression Analysis
To correct for confounding, the 23 significant univariate predictors were analyzed by multivariate logistic regression. This analysis yielded five independent predictors of hospital mortality (Table 6 ): (1) the hypoxemia index; (2) the number of nonpulmonary organ systems failed; (3) immunosuppression; (4) age > 80 years; and (5) preexisting medical prognosis for survival < 5 years. The Hosmer-Lemeshow {chi}2 goodness-of-fit statistic had a p value of 0.91, indicating excellent model fit. Based on the magnitude of the Wald statistic, the hypoxemia index was the strongest independent predictor.


View this table:
[in this window]
[in a new window]

  		Table 6.. Multivariate Logistic Regression Analysis*

 
A mortality prediction rule was formulated by summing the products of each independent predictor with its coefficient in the logistic regression equation. This calculation yields a score given by the following equation:

(2)
We classified predicted outcome into three categories based on the score: (1) score < 9.20 predicts hospital survival, with mortality risk < 50%; (2) score between 9.20 and 12.14 predicts mortality with risk 50% to 95%; and (3) score > 12.14 predicts mortality with > 95% risk.

At a 50% mortality threshold (score = 9.20), the model correctly classified outcome in 88% of cases, with a sensitivity of 85%, specificity of 90%, positive predictive value of 85%, and negative predictive value of 90%. A score > 12.14 identifies patients with an extremely high mortality risk (> 95% probability of death); 28 subjects had a score > 12.14. Of note, no survivors were misclassified at the 95% mortality threshold, yielding a positive predictive value of 100% and specificity of 100% (lower bound of 95% CI, 89%).

ROC curves assess outcome discrimination, ie, death vs survival, and are based on sensitivity and specificity at all thresholds.24 The area under the ROC curve (AUC) is an indicator of test performance; an AUC of 1 indicates perfect performance, whereas an AUC of 0.5 indicates that a test performs as well as chance, ie, "fifty-fifty" discrimination. ROC analysis yielded an AUC for our prediction model of 0.95 (95% CI, 0.91 to 0.99), which was significantly better than those for SAPS and APACHE II, at 0.69 (95% CI, 0.59 to 0.78) and 0.76 (95% CI, 0.67 to 0.85), respectively (Fig 1 ).



View larger version (34K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 1.. ROC comparison.

 
Cross-Validation
This method uses the variability in the original data sample to simulate prediction performance in a new population.23 One subject is removed from the original data set, and then significant predictors are determined by univariate analysis. Stepwise logistic regression is then performed to derive a new prediction model. The new equation is used to calculate the probability of hospital mortality for the left-out case. The procedure is repeated for all remaining cases.

The five variables selected by our original prediction model (hypoxemia index, nonpulmonary organ system failure, immunosuppression, age > 80 years, and medical prognosis < 5 years) were selected in 89% of the cross-validation trials. At a mortality threshold of 50%, cross-validation misclassified 17% of cases, compared with the misclassification rate of 12% in the original data set. This predicts a classification error rate approaching 17% when our rule is applied to other populations. At a 95% mortality threshold, the positive predictive value and specificity each remained 100% in the cross-validated model, ie, no survivors were misclassified at the 95% mortality threshold. The AUC for the cross-validated model was 0.90 (95% CI, 0.84 to 0.96; Fig 1 ).

Comparison of Lung Injury Indices in Prediction Rules
Table 7 displays the performance characteristics of nine different prediction models; each has one measure of lung injury forced into the model (hypoxemia index, PaO2/PAO2, PaO2/FIO2, lung injury score, presence of ARDS, P(A-a)O2, FIO2, or PEEP), along with the nonpulmonary variables selected by stepwise logistic regression. In addition, a model is shown that has no measure of lung injury included. Correct classification rates, AUCs, the number of high-risk patients identified (> 95% predicted mortality), and goodness-of-fit tests are compared. In all categories (except for model fit), the model that utilized the hypoxemia index yielded the best performance, although the other models performed nearly as well. Of note, at a 95% mortality threshold, the positive predictive value and specificity were each 100% for all nine models, ie, no survivors were misclassified at this threshold.


View this table:
[in this window]
[in a new window]

  		Table 7.. Prediction Model Comparison

 

   Discussion
TOP
 Abstract
 Introduction
 Materials and Methods
 Results
 Discussion
References
 
We have found that the extent of lung injury as well as other nonpulmonary factors are important independent markers of prognosis in patients with CAP requiring mechanical ventilation. These predictors were incorporated into a model that shows good discriminative ability (88% accuracy in outcome classification) in our study. Furthermore, the model identified patients with the poorest prognosis (> 95% probability of death in hospital) without misclassifying survivors. Importantly, the model utilizes data readily obtained within the first 24 h of instituting mechanical ventilation. Based on ROC analysis, the prediction rule’s performance was superior to that of SAPS and APACHE II in our population of patients with CAP and respiratory failure.

This study has some limitations. Because the rule was derived from a relatively small population of less than 150 patients, the regression coefficients might vary if the derivation were based on a larger group. There was discrepancy between the error rate via the cross-validated model (17%) and the original error rate (12%), suggesting somewhat poorer outcome discrimination (at a 50% mortality threshold) when applied to other populations. However, the 95% mortality threshold retained 100% positive predictive value and specificity, even in the cross-validated model. Observed mortality may have been underestimated compared with the larger population of subjects with CAP and respiratory failure: some patients may have been systematically excluded from mechanical ventilation because of clinicians’ perception of a low chance of survival or patients’ informed refusal of life-sustaining treatment relating to personal preference or perceived poor prognosis. Our population was from a tertiary-care university hospital, which may limit generalizability to other institutions. Finally, the prediction rule requires further validation in a separate cohort before it can be used reliably in a clinical setting. Specifically, the 95% mortality threshold (score > 12.14) must be shown to be robust in a much larger number of patients before it can serve as an indication for withdrawal of intensive care.

There was no standardized threshold for instituting mechanical ventilation. We relied on the clinicians’ judgment at the time of illness to determine whether mechanical ventilation was appropriate. Therefore, mechanical ventilation was initiated at various levels of respiratory failure ranging from moderate to severe. This variation has, however, allowed us to quantify the contribution of a wide range of lung injury to mortality risk. For example, the observed values for our patients’ hypoxemia indices ranged from 10 to 94 (on a scale from 0 to 100). Across this range, mortality varies markedly. To illustrate, a hypothetical patient who is older than 80 years, is immunosuppressed, has no nonpulmonary organ systems failed, and does not have a medical prognosis < 5 years has a predicted hospital mortality of 32% at a hypoxemia index of 40, but the risk increases to 96% at a hypoxemia index of 90.

This is the first study to report that the degree of lung injury is the most important prognostic criterion in mechanically ventilated patients with CAP. Only a few studies of severe CAP in the ICU have incorporated multivariate analysis.2 3 7 Prognostic factors reported by these studies were as follows: anticipated death within 5 years, shock, bacteremia, nonpneumonia-related complications, ineffective antibiotic therapy, radiographic spread of pneumonia, SAPS > 13, and infection by S pneumoniae or Enterobacteriaceae. There are several possible reasons why the predictors in our study differ. We studied only mechanically ventilated patients, whereas all prior studies of severe CAP had mixed populations of ventilated and nonventilated patients. Thus, as one would expect, lung injury becomes a much more important variable in our population and may have outweighed other variables that were significant in prior studies, such as bacteremia, radiographic spread of pneumonia, and SAPS score. Secondly, because the major focus of our paper was to establish early prognostic estimates, we did not examine variables such as response to antibiotics and nonpneumonia-related complications that would require collection of data beyond the first day of mechanical ventilation. We did find, however, that none of the antibiotics in Table 3 were significantly associated with survival in the multivariate analysis. This suggests either that antibiotic selection did not differ significantly between the two groups, or that factors other than antibiotic selection have the strongest impact on outcome in the mechanically ventilated patient. Finally, of those published studies of severe CAP that utilized multivariate analysis, observed mortality ranged from 22% to 28%.2 3 7 The observed mortality in our study was much higher (46%), arguing that our population consisted of sicker patients with different markers of mortality. The higher mortality, and thus greater number of end points, in our study lends greater strength to the selection of predictor variables in the analysis.

Because the degree of acute lung injury may be a dynamic process in CAP, especially during the first 24 h of mechanical ventilation when pneumonia is in evolution and antibiotics are taking effect, we speculated a priori that isolated single measurements in time (ie, P(A-a)O2, PaO2/PAO2 ratio, PaO2/FIO2 ratio, PEEP, and FIO2) may not accurately reflect the state of lung injury; therefore, we derived the hypoxemia index to account for fluctuations and to summarize the state of derangement throughout a 24-h interval. When compared with the other correlates of lung injury (P(A-a)O2, PaO2/PAO2 ratio, PaO2/FIO2 ratio, PEEP, FIO2, presence of ARDS, and the lung injury score), logistic regression selected the hypoxemia index as the best predictor of risk. Furthermore, direct comparison of separate logistic regression models, each employing one of these eight measures of lung injury in addition to the nonrespiratory predictors, showed that use of the hypoxemia index led to the best model performance. However, as shown in Table 6 , the models using other measures of lung injury performed nearly as well, implying that lung injury can be measured by several methods.

There is one significant limitation to the use of the hypoxemia index. If arterial blood gases are not drawn frequently, the true minimal FIO2 to maintain PO2 > 60 mm Hg will not be known and the value for the hypoxemia index may be inaccurate. In such cases, it may be prudent to record the minimum FIO2 to maintain the pulse oximeter arterial oxygen saturation at > 90% as a surrogate measurement. Furthermore, it remains to be determined whether the hypoxemia index retains good predictive ability in other ICU cohorts with acute respiratory failure, ie, COPD, ARDS, bone marrow transplant recipients, and AIDS with P carinii pneumonia. A prospective study is needed to further compare the performance of the hypoxemia index with standard measurements of lung injury.

The estimated mean mortality by our prediction rule was relatively low (mean, 29%) among 11 patients who survived an initial trial of mechanical ventilation, were subsequently given a do not resuscitate/do not intubate (DNR/DNI) order, then died of recurrent respiratory failure when reintubation was withheld. Reasons for the DNR/DNI orders were mixed and included patient preference for less aggressive care, patient/clinician estimate of poor quality of life if survival was achieved, and clinician’s overall estimate of survival being extremely low. Because of the retrospective nature of this study, the primary reason for the DNR/DNI orders could not be determined by chart review for some patients. These results raise the possibility, however, that clinicians may have overestimated mortality risk in some cases. Indeed, studies have shown there is significant variability in physicians’ estimates of survival among critically ill patients,25 26 27 arguing for more widespread use of quantitative prediction rules that can reduce the level of prognostic uncertainty.

The better performance of our prognostic model in comparison to APACHE II and SAPS (as shown by ROC analysis) suggests that disease-specific prognostic indices might be more accurate for evaluating uniform intensive care populations than indices derived from heterogeneous groups (APACHE II, SAPS). Such indices utilize physiologic parameters without weighting of organ system derangements according to the predominant underlying pathology (ie, lung injury in pneumonia). Indeed, other investigators have found that generic indices such as APACHE II do not perform well when applied to homogeneous populations consisting of AIDS patients, cardiac arrest survivors, and acute renal failure patients.28 29 30

In summary, we have derived a prediction rule for estimating mortality in patients with CAP requiring mechanical ventilation, utilizing data obtained during the first 24 h of assisted ventilation. Among the predictor variables, the extent of lung injury measured by the hypoxemia index is the most important prognostic factor. In this population, the model correctly classified outcome in 88% of cases and identified high-risk patients (> 95% predicted mortality) with 100% positive predictive value and 100% specificity. Although the model requires further validation, it may be useful as an objective, quantitative aid for identifying patients with far-advanced pneumonia and respiratory failure whose condition may be irreversible despite aggressive intensive care.


   Footnotes
 
Abbreviations: APACHE II = acute physiology and chronic health evaluation II; AUC = area under the curve; CAP = community-acquired pneumonia; CI = confidence interval; DNR/DNI = do not resuscitate/do not intubate; FIO2 = fraction of inspired oxygen; P(A-a)O2 = alveolar-arterial oxygen pressure difference; PAO2 = alveolar partial pressure of oxygen; PEEP = positive end-expiratory pressure; ROC = receiver operating characteristics; SAPS = Simplified Acute Physiology Score; E = minute ventilation

Received for publication March 24, 1999. Accepted for publication August 11, 1999.


   References
TOP
 Abstract
 Introduction
 Materials and Methods
 Results
 Discussion
 References
 

  1. Leeper, K (1996) Severe community-acquired pneumonia. Semin Respir Infect 11,96-108[Medline]
  2. Torres, A, Serra-Batlles, J, Ferrer, A, et al (1991) Severe community-acquired pneumonia: epidemiology and prognostic factors. Am Rev Respir Dis 144,312-318[ISI][Medline]
  3. Moine, P, Vercken, J, Chevret, S, et al (1994) Severe community-acquired pneumonia: etiology, epidemiology, and prognostic factors. Chest 105,1487-1495[Abstract/Free Full Text]
  4. Marrie, TJ, Durant, H, Yates, L (1989) Community-acquired pneumonia requiring hospitalization: 5-year prospective study. Rev Infect Dis 11,586-599[ISI][Medline]
  5. Sorensen, J, Cederholm, I, Carlsson, C (1986) Pneumonia: a deadly disease despite intensive care treatment. Scand J Infect Dis 18,329-335[ISI][Medline]
  6. Almirall, J, Mesalles, E, Klamgurg, J, et al (1995) Prognostic factors of pneumonia requiring admission to the intensive care unit. Chest 107,511-516[Abstract/Free Full Text]
  7. Leroy, O, Santre, C, Beuscart, C, et al (1995) A five-year study of severe community-acquired pneumonia with emphasis on prognosis in patients admitted to an intensive care unit. Intensive Care Med 21,24-31[CrossRef][ISI][Medline]
  8. Leroy, O, Georges, H, Beuscart, C, et al (1996) Severe community-acquired pneumonia in ICUs: prospective validation of a prognostic score. Intensive Care Med 22,1307-1314[ISI][Medline]
  9. Enger, C, Graham, N, Peng, Y, et al (1996) Survival from early, intermediate, and late stages of HIV infection. JAMA 275,1329-1334[Abstract]
  10. Salerno, F, Borroni, G, Moser, P, et al (1993) Survival and prognostic factors of cirrhotic patients with ascites: a study of 134 outpatients. Am J Gastroenterol 88,514-519[ISI][Medline]
  11. Strom, K (1993) Survival of patients with chronic obstructive pulmonary disease receiving long-term domiciliary oxygen therapy. Am Rev Respir Dis 147,585-591[ISI][Medline]
  12. Smith, W (1985) Epidemiology of congestive heart failure. Am J Cardiol 55,3A-8A[CrossRef][Medline]
  13. Mailloux, L, Bellucci, A, Napolitano, B, et al (1994) Survival estimates for 683 patients starting dialysis from 1970 through 1989: identification of risk factors for survival. Clin Nephrol 42,127-135[ISI][Medline]
  14. Alexanian, R, Dimopoulos, M (1994) The treatment of multiple myeloma. N Engl J Med 330,484-489[Free Full Text]
  15. Greenberg, P, Cox, C, Lebeaux, M, et al (1997) International scoring system for evaluating prognosis in myelodysplastic syndromes. Blood 89,2079-2088[Abstract/Free Full Text]
  16. Buchner, T, Heinecke, A (1996) The role of prognostic factors in acute myelogenous leukemia. Leukemia 10(Suppl 1),S28-S29
  17. Murray, JF, Matthay, M, Luce, J, et al (1988) An expanded definition of the adult respiratory distress syndrome. Am Rev Respir Dis 138,720-723[ISI][Medline]
  18. Le Gall, JR, Loirat, P, Alperovitch, A, et al (1984) A simplified acute physiology score for ICU patients. Crit Care Med 12,975-977[ISI][Medline]
  19. Knaus, W, Draper, E, Wagner, D, et al (1985) APACHE II: a severity of disease classification system. Crit Care Med 13,818-829[ISI][Medline]
  20. Bernard, GR, Artigas, A, Brigham, KL, et al (1994) The American-European Consensus Conference on ARDS: definitions, mechanisms, relevant outcomes, and clinical trial coordination. Am J Respir Crit Care Med 149,818-824[Abstract]
  21. Knaus, W, Draper, E, Wagner, D, et al (1985) Prognosis in acute organ-system failure. Ann Surg 202,685-693[ISI][Medline]
  22. Hosmer, D, Lemeshow, S (1989) Applied logistic regression. ,140-145 Wiley New York, NY.
  23. Efron, B, Gong, G (1983) A leisurely look at the bootstrap, the jacknife, and cross-validation. Am Statistician 37,36-48
  24. Beck, J, Shultz, E (1986) The use of relative operating characteristic (ROC) curves in test performance evaluation. Arch Pathol Lab Med 110,13-20[ISI][Medline]
  25. Pearlman, R (1987) Variability in physician estimates of survival for acute respiratory failure in chronic obstructive pulmonary disease. Chest 91,515-521[Abstract/Free Full Text]
  26. Perkins, H, Jonsen, A, Epstein, W (1986) Providers as predictors: using outcome predictions in intensive care. Crit Care Med 14,105-110[ISI][Medline]
  27. Poses, R, Bekes, C, Copare, F, et al (1989) The answer to "what are my chances, doctor?" depends on whom is asked: prognostic disagreement and inaccuracy for critically ill patients. Crit Care Med 17,827-833[ISI][Medline]
  28. Brown, M, Crede, W (1995) Predictive ability of APACHE II scoring applied to HIV-positive patients. Crit Care Med 23,848-853[CrossRef][ISI][Medline]
  29. Niskanen, M, Kari, A, Nikki, P, et al (1991) APACHE II and Glasgow coma scores as predictors of outcome from intensive care after cardiac arrest. Crit Care Med 19,1465-1473[Medline]
  30. Schaefer, J, Jochimsen, F, Keller, F, et al (1991) Outcome prediction of acute renal failure in medical intensive care. Intensive Care Med 17,19-24[CrossRef][ISI][Medline]



This article has been cited by other articles:


Home page
 ChestHome page
R. Mehta, M. Groth, and F. Pascual
Clinical Application of a Prognostic Model for Severe Community-Acquired Pneumonia
Chest, January 1, 2001; 119(1): 312 - 313.
[Full Text] [PDF]

Home page
 ChestHome page
C. E. Behrendt
Acute Respiratory Failure in the United States : Incidence and 31-Day Survival
Chest, October 1, 2000; 118(4): 1100 - 1105.
[Abstract] [Full Text] [PDF]

This Article
Right arrow Abstract Freely available
Right arrow Full Text (PDF) Free
Right arrow Submit a response
Right arrow Alert me when this article is cited
Right arrow Alert me when eLetters are posted
Right arrow Alert me if a correction is posted
Right arrow Citation Map
Services
Right arrow Email this article to a friend
Right arrow Similar articles in this journal
Right arrow Similar articles in ISI Web of Science
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Add to My Personal Article Archive
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via ISI Web of Science (18)
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Pascual, F. E.
Right arrow Articles by Wachter, R. M.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Pascual, F. E.
Right arrow Articles by Wachter, R. M.

HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS