Introduction

The ability of noninvasive ventilation (NIV) to reduce intubation and mortality has been clearly established in patients with acute-on-chronic respiratory failure (AOC) [12345] or to reduce intubation and complications for cardiogenic pulmonary edema (CPE) [678]. In patients with acute de novo respiratory failure (de novo), i.e., not that exacerbating a chronic lung or cardiac insufficiency, randomized controlled trials showed reduced mortality with NIV in highly selected populations such as patients with immunosuppression [9] or specific postoperative conditions [1011]. Although several studies have demonstrated the ability of this technique to reduce endotracheal intubation (ETI) in highly selected patients [12131415], there are no multicenter studies showing that NIV reduces mortality in large populations of patients with de novo respiratory failure due to a broad array of causes excluding acute pulmonary edema [1516]. By contrast, observational studies in patients with community-acquired pneumonia, i.e., probably with less strict selection criteria, found high NIV failure rates [171819]. In addition, a number of studies suggested that NIV may be harmful at least in specific circumstances or indications by delaying intubation [202122]. This important question has not yet been resolved.

We therefore examined whether the risks and benefits of NIV differ between patients with de novo and those with CPE or AOC. To investigate this we took the opportunity to analyze data from a large survey on mechanical ventilation performed in 2002 and presented in a companion paper [23]. Patients included in the survey were divided into a group with de novo and a group with either CPE or AOC. Intensive care unit (ICU) mortality, incidence of nosocomial pneumonia, and ICU length of stay were compared in these two groups after adjustment for disease severity.

Methods

Participating centers and patients

A prospective observational study was conducted in 70 French ICUs between 4 and 24 March 2002. All ICU patients who required ventilatory assistance for acute respiratory failure at any time during the ICU stay were included. The investigators completed a questionnaire for each patient to collect data from admission to ICU discharge or death. The major characteristics of the participating centers and patients are presented in the companion paper [23]. During the 3-week study period 1,943 patients were admitted to the participating ICUs, 1,076 of whom received ventilatory support. Because the study sought to compare patients with ETI vs. those with NIV, we restricted our analysis to the following group [23], we excluded 21 patients who were already tracheostomized before ICU admission, 358 ventilated for coma, and 130 following surgery as NIV is not widely used in these two situations [23]. Finally, because a main end-point was patient outcome, we also excluded the 35 patients treated with NIV who had a do-not-intubate order, and 8 who were discharged in another ICU still intubated with no data on final ICU outcome. This left 524 patients for the analysis: 299 in the de novo group and 225 in the CPE-AOC group, including 152 patients with AOC and 73 with pulmonary edema.

Data collection

Patients' data included age, gender, location before hospital and ICU admission, Simplified Acute Physiology Score (SAPS) II, Logistic Organ Dysfunction (LOD) score [2425], and severity of underlying disease [26]. We recorded data on immunosuppression (defined as neutropenia below 1000/mm3 after bone marrow transplantation or anticancer chemotherapy, immunosuppressive therapy for solid organ transplantation, or connective tissue disorder requiring corticosteroid therapy of at least 20 mg/day for at least 3 weeks). The ventilation modality was recorded as NIV or ventilation via ETI. In patients given NIV as first-line ventilatory support, efficacy was recorded as successful NIV (no need for ETI) or failed NIV (secondary need for ETI or death before ETI). ICU mortality, nosocomial pneumonia, ICU length of stay, and length of mechanical ventilation (NIV and conventional mechanical ventilation) were recorded.

Data quality

All patient forms were reviewed by the principal investigator. Investigators were contacted about errors and missing data. A double data-entry procedure was used, and data that triggered the alert system or showed discrepancies were checked.

Statistical analysis

The median and interquartile range (IQR) are reported for continuous variables and absolute and relative frequencies for categorical variables as well as the 95% confidence intervals (95% CI) where appropriate. Each potential risk factor for death, nosocomial pneumonia, longer length of mechanical ventilation, or longer ICU length of stay was evaluated in a univariate model. For continuous data we used Student's t test or the Mann-Whitney U test as appropriate and for categorical variables the χ2 test or Fisher's exact test. All factors yielding p values less than 0.25 were entered into the logistic regression models. We separated the population in two major groups of interest (the de novo group and the CPE-AOC group), and eight logistic regression models were first constructed (four different outcomes in these two different populations). Continuous variables were dichotomized according to the median in the population. In these multivariate analyses ventilatory support was entered as a single variable with three distinct categories: invasive mechanical ventilation was the reference category with an adjusted odds ratio (OR) at 1.00 and was compared to NIV success and NIV failure. All selected variables were entered in the models and were examined for collinearity (e.g., the LOD score was not included because of collinearity with the SAPS II score). Three other logistic regression models were also performed, one in the NIV failure group, and finally in the de novo group and in the CPE-AOC group looking at the use of NIV and its association with outcome. Goodness-of-fit was assessed using the Hosmer–Lemeshow χ2 test, and discrimination using the area under the receiver operating characteristic (ROC) curve for each model. Accuracy was considered good when the value ranged from 0.70 to 0.80 and excellent when it was greater than 0.80. The adjusted OR and 95% confidence interval (CI) were calculated for each factor that was statistically significant in the logistic regression model. Any p values less than 0.05 were considered significant. All tests were two-tailed. Statistical tests were performed using Intercooled STATA 8.2 software (StatCorp LP, Texas, USA).

Results

Mortality

Overall mortality was 32% (95% CI, 28–37%). Mortality was significantly lower in the NIV group, i.e., patients receiving NIV as first-line treatment, than in the ETI group, i.e., patients receiving ETI as first-line treatment (23%, 95% CI 18–30%, vs. 39%, 95% CI, 33–44%; p < 0.0001), and was higher in the de novo group than in the CPE-AOC group (Table 1).

Table 1 Comparison of patients with de novo respiratory failure and of patients with cardiogenic pulmonary edema or acute-on-chronic respiratory failure: median (and interquartile range) or absolute number (and percentage) (de novo de novo acute respiratory failure, CPE-AOC cardiogenic pulmonary edema or acute-on-chronic respiratory failure, FIO 2 fraction of inspired oxygen, LOD Logistic Organ Dysfunction system, NIV noninvasive ventilation, PaCO 2 arterial carbon dioxide tension, PaO 2 arterial oxygen dioxide tension, SAPS Simplified Acute Physiology Score, VAP ventilator-associated pneumonia, MV mechanical ventilation)

In the de novo group factors significantly associated with death in the univariate analysis were older age, worse SAPS II, ultimately or rapidly fatal disease, and immunosuppression (Table 2). The NIV success rate was higher in survivors, and the NIV failure rate was higher in nonsurvivors. In the logistic regression analysis successful NIV was independently associated with survival, whereas NIV failure, worse SAPS II, and immunosuppression were independently associated with death (Fig. 1).

Fig. 1
figure 1

Independent risk factors for death in patients with acute de novo respiratory failure (de novo, black squares) and in those with cardiogenic pulmonary edema or acute-on-chronic respiratory failure (CPE-AOC, white squares). In all multivariate analyses presented in the figures ventilatory support was entered as a single variable with three distinct categories: invasive mechanical ventilation was the reference category with an OR at 1.00 and was compared to NIV success and NIV failure. NA Not available (variable not entered in the logistic regression model); NIV noninvasive ventilation; SAPS Simplified Acute Physiology Score; * age above 64 years in patients with acute de novo respiratory failure and 72 years in those with cardiogenic pulmonary edema or acute-on-chronic respiratory failure; ** SAPS II above 47 points in patients with acute de novo respiratory failure and 38 points in those with cardiogenic pulmonary edema or acute-on-chronic respiratory failure; § presence of an ultimately or rapidly fatal disease

Table 2 Potential risk factors for death, longer ICU stay, longer length of mechanical ventilation, and nosocomial pneumonia in patients with de novo acute respiratory failure (n = 299): percentage (and 95% confidence interval) (ICU intensive care unit, LOS length of stay, LOMV length of mechanical ventilation, NIV noninvasive ventilation, SAPS Simplified Acute Physiologicy Score, MV mechanical ventilation)

In the CPE-AOC group the univariate analysis identified worse SAPS II, ultimately or rapidly fatal underlying disease, and nosocomial pneumonia as significantly associated with death (Table 3). The NIV success rate was higher in survivors, but the failure rate did not differ between survivors and nonsurvivors. In the logistic regression analysis NIV failure had no independent association with death, but successful NIV remained independently associated with survival. A worse SAPS II and an ultimately or rapidly fatal disease were independently associated with death (Fig. 1).

Table 3 Potential risk factors for death, longer ICU stay, longer length of mechanical ventilation, and nosocomial pneumonia in patients with cardiogenic pulmonary edema or acute-on-chronic respiratory failure (n = 225): percentage (and 95% confidence interval) (ICU intensive care unit, LOS length of stay, LOMV length of mechanical ventilation, NIV noninvasive ventilation, SAPS Simplified Acute Physiology Score, MV mechanical ventilation)

Nosocomial pneumonia

Sixty-seven patients (13%, 95% CI, 10–16%) experienced at least one episode of nosocomial pneumonia. Nosocomial pneumonia rates were not statistically different in the NIV and ETI groups (10%, 95% CI 7–15%, vs. 14%, 11–19%; p = 0.16). The causative pathogen was identified in protected samples in 55 of the 56 cases (24 bronchoalveolar lavage, 22 plugged telescopic catheter, 9 protected specimen brush) and in unprotected samples in 12 cases. In the NIV group 7 patients acquired pneumonia while receiving NIV and 15 while receiving endotracheal ventilation after NIV failure. The nosocomial pneumonia rate was higher in the de novo group than in the CPE-AOC group (16% vs. 8%, p = 0.01; Table 1). In both groups type of ventilatory support was the only factor significantly associated with the risk of nosocomial pneumonia in the univariate analysis (Tables 23). In the logistic regression analysis successful NIV was the only independent factor associated with absence of nosocomial pneumonia in both the de novo and CPE-AOC groups. NIV failure was not associated with a higher risk of pneumonia (Fig. 2).

Fig. 2
figure 2

Independent risk factors for nosocomial pneumonia in patients with acute de novo respiratory failure (de novo, black squares) and in those with cardiogenic pulmonary edema or acute-on-chronic respiratory failure (CPE-AOC, white squares). For details see legend of Fig. 1. NA Not available (variable not entered in the logistic regression model); NIV noninvasive ventilation; SAPS Simplified Acute Physiology Score; * age above 64 years in patients with acute de novo respiratory failure and 72 years in those with cardiogenic pulmonary edema or acute-on-chronic respiratory failure; ** SAPS II above 47 points in patients with acute de novo respiratory failure and 38 points in those with cardiogenic pulmonary edema or acute-on-chronic respiratory failure

ICU stay length

As shown in Table 1, the ICU stay was significantly shorter in the NIV than in the ETI group (7 days, IQR 4–15 days, vs. 10 days, 5–20 days; p = 0.03). The de novo patients spent more time in the ICU than the CPE-AOC patients (10 days, 5–21 days, vs. 7 days, 4–13 days; p = 0.0001). In the de novo group worse SAPS II and occurrence of nosocomial pneumonia were independently associated with a longer stay, whereas successful NIV was associated with a shorter stay (Table 2, Fig. 3). NIV failure did not predict longer ICU stay. In the CPE-AOC group occurrence of nosocomial pneumonia and NIV failure were independent risk factors for a longer stay, whereas successful NIV and younger age were significantly associated with a shorter stay (Table 3). The logistic regression analysis confirmed these results (Fig. 3). Entering length of mechanical ventilation as the dependent variable in the logistic regression models produced nearly identical results (Fig. 4).

Fig. 3
figure 3

Independent risk factors for longer ICU length of stay (LOS) in patients (higher than the median LOS) with acute de novo respiratory failure (de novo, black squares) and in those with cardiogenic pulmonary edema or acute-on-chronic respiratory failure (CPE-AOC, white squares). For details see legend of Fig. 1. NA Not available (variable not entered in the logistic regression model); NIV noninvasive ventilation; SAPS simplified acute physiology score; *age above 64 years in patients with acute de novo respiratory failure and 72 years in those with cardiogenic pulmonary edema or acute-on-chronic respiratory failure; ** SAPS II above 47 points in patients with acute de novo respiratory failure and 38 points in those with cardiogenic pulmonary edema or acute-on-chronic respiratory failure; § presence of an ultimately or rapidly fatal disease;  ICU length of stay 10 days or longer in patients with acute de novo respiratory failure and 7 days in those with cardiogenic pulmonary edema or acute-on-chronic respiratory failure

Fig. 4
figure 4

Independent risk factors for longer length of mechanical ventilation (LOMV) in patients (higher than the median LOMV) with acute de novo respiratory failure (de novo, black squares) and in those with cardiogenic pulmonary edema or acute-on-chronic respiratory failure (CPE-AOC, white squares). For details see legend of Fig. 1. NA Not available (variable not entered in the logistic regression model); NIV noninvasive ventilation; SAPS simplified acute physiology score; * age above 64 years in patients with acute de novo respiratory failure and 72 years in those with cardiogenic pulmonary edema or acute-on-chronic respiratory failure; ** SAPS II above 47 points in patients with acute de novo respiratory failure and 38 points in those with cardiogenic pulmonary edema or acute-on-chronic respiratory failure;  ICU length of stay 10 days or longer in patients with acute de novo respiratory failure and 7 days in those with cardiogenic pulmonary edema or acute-on-chronic respiratory failure

Balance between risks from failure and benefits from success

The last analysis looked at the overall effect of using NIV (taken as an independent variable) on outcome in the two groups. In the CPE-AOC group, the use of NIV was significantly associated with a better outcome, indicating a positive balance in favor of the use of NIV (adjusted OR 0.33, 95% CI 0.15–0.73; p = 0.007). In the de novo group, the use of NIV was not significantly associated with a better or worse outcome, indicating that the balance was neutral (adjusted OR 1.18, 95% CI 0.66–2.10; p = 0.569).

Discussion

The results of the present study show that NIV, when successful, was independently associated with survival, a decreased risk of nosocomial pneumonia in the CPE-AOC group, and a shorter ICU stay than with ETI used as the standard treatment for acute respiratory failure in both de novo and CPE-AOC patients. NIV failure was independently associated with death in the de novo group but not in the CPE-AOC group, and with longer durations of mechanical ventilation and ICU stay in the CPE-AOC group but not the de novo group, and was not associated with a higher rate of nosocomial pneumonia.

The 2002 survey on mechanical ventilation showed a significant increase in NIV use in French ICUs [23] since 1997 [27]. In the multivariate analysis good NIV tolerance and a high body mass index were associated with successful NIV, whereas a worse SAPS II or a diagnosis of de novo respiratory failure was associated with NIV failure and subsequent ETI. The term “de novo” is used here to designate pneumonia, gastric content aspiration, atelectasis, and pulmonary, and extrapulmonary acute respiratory distress syndrome in patients free of chronic lung disease. As opposed to de novo, the two other common patterns of acute respiratory failure, namely, CPE and AOC, were associated with successful NIV [27]. Thus the cause of respiratory failure affects the likelihood that NIV will be successful. However, the main goal is not to achieve successful NIV but rather to reduce mortality and morbidity, as with all new treatments. Consequently in the present study we assessed the interaction between diagnosis and NIV use on mortality, length of stays, and nosocomial pneumonia.

Impact of NIV on outcome

In both the de novo and the CPE-AOC groups independent risk factors associated with death in the multivariate model were a higher SAPS II, an ultimately or rapidly fatal underlying disease, and immunosuppression, whereas successful NIV was independently associated with survival. Successful NIV has been found beneficial in clinical trials in various populations and was also an independent predictor of survival in the 1997 French survey on mechanical ventilation [27], establishing that the benefits observed in patients selected to clinical trials occurred also in the unselected patients seen in everyday practice. Moreover, the dramatic increase in NIV use shown in the 2002 French survey [23] combined with the unchanged NIV success rate translated into an increase in the absolute number of patients successfully treated with NIV. A worse SAPS II at admission, an ultimately or rapidly fatal disease, and immunosuppression independently predicted mortality, in keeping with findings from the 1997 French survey on mechanical ventilation [27].

In the present study the impact of NIV on ICU mortality differed between the two groups, and this is explained by the different impact of NIV failure: NIV failure was independently associated with ICU mortality in the de novo group but not in the CPE-AOC group. This association may reflect either a causal effect (with NIV failure contributing to cause death) or a confounding effect (with NIV failure merely being a marker for more severe disease). Our analysis incorporated important prognostic scores, age, and indicators of underlying disease to adjust on severity, making less likely that this association was simply a marker of severity. In addition, the different results obtained with the same analysis in the other group, where the most severe patients are also expected to be found in the subgroup of NIV failure, suggests that there is a real difference in the effect of NIV failure per se between the two subgroups. Although data from studies of NIV have suggested potential harmful effects of NIV in patients [142122], this important finding has not been reported previously. In a randomized controlled trial excluding patients with chronic respiratory failure Wysocki et al. [14] found no significant reduction in ICU mortality in the NIV group (33% vs. 50% in the control group); however, in the subgroup with acute hypoxemic nonhypercapnic respiratory failure a trend to a higher mortality existed among patients treated with NIV. In de novo patients, a beneficial effect of NIV on mortality has been found chiefly in specific patient subgroups such as patients with immunosuppression [9], solid organ transplantation [10], or lung resection [10] or in very selected patients [13]. In the only multicenter study to date in patients with acute hypoxemic nonhypercapnic respiratory failure, Delclaux et al. [20] found that continuous positive airway pressure was associated with a higher adverse event rate (cardiac arrest and gastrointestinal bleeding) and failed to improve outcomes compared to oxygen alone. In a recently published multicenter trial evaluating the impact of NIV in postextubation respiratory failure, mortality was higher with NIV than with standard medical therapy [21].

In the present study the patients with CPE or AOC benefited the most from NIV. This finding is consistent with previous studies. In patients with AOC relatively large randomized controlled trials showed reductions in ETI and mortality rates [34]. In a mixed population of patients with CPE or AOC, Girou et al. [728] recently found that ICU mortality was lowered thanks to an increasing use of NIV replacing ETI. Lastly, two studies comparing patients with chronic obstructive pulmonary disease treated with NIV or endotracheal intubation at a very late stage of respiratory acidosis did not find any evidence of worse outcome despite a high rate of failure [2930]. In addition, the outcome of NIV failure in this subgroup seemed superimposable on that of patients receiving endotracheal intubation at onset [30].

Nosocomial pneumonia

The rate of nosocomial pneumonia was lower in the NIV group, but this difference was not statistically significant. The rate of nosocomial pneumonia was lower in the CPE-AOC than in the de novo group. In both groups successful NIV was the only factor independently associated with a lower nosocomial pneumonia rate. This protective effect of NIV has been reported in single [28] and multicenter [17] studies and can be attributed both to the absence of an endotracheal tube bypassing the upper airways and to decreased invasiveness of the overall management. Nosocomial pneumonia is associated with increased in morbidity and mortality in ICU patients [31]. In the present study ETI after NIV failure did not lead to a higher nosocomial pneumonia rate compared to first-line ETI.

Length of stay

The median length of ICU stay was significantly lower in the NIV group (7 days, IQR 4–15) than in the ETI group (10 days, 5–20; p = 0.003) and was significantly lower in the CPE-AOC group (7 days, 4–13) than in the de novo group (10 days, 5–21; p = 0.0001). Similar patterns were observed for the duration of mechanical ventilation. In both groups nosocomial pneumonia was independently associated with a longer duration of mechanical ventilation and therefore a longer ICU stay, whereas successful NIV was independently associated with a shorter ICU stay. NIV has not consistently been shown to affect ICU length of stay in randomized control trials. Two trials found a reduction in ICU length of stay with NIV in de novo patients [932]. In AOC patients NIV reduced ICU length of stay in one randomized controlled trial and one retrospective matched case-control study [328]. Our findings suggest that, compared to first-line ETI, successful NIV may reduce the ICU length of stay. NIV failure was not associated with a longer length of ICU stay in the de novo patients. Although NIV failure was associated with increased length of ICU stay in the CPE-AOC patients, there was no increase in mortality, indicating that failed NIV followed by timely ETI involved no loss of chance of survival in this population.

Clinical consequences

NIV failure followed by ETI was independently associated with ICU mortality in de novo patients but not CPE-AOC patients. This explained the different impact of the use of NIV in the two populations of patients. As discussed above, this study does not provide a definitive explanation for this association, but there is the possibility that in de novo patients NIV failure is associated with loss of chance. The interest of this study is also to represent real life and not the context of a randomized controlled trial with carefully selected patients. This may have important clinical implications because one plausible explanation is that inadequate or prolonged use of NIV delays ETI in de novo patients, thereby jeopardizing the patient's chance of survival. Criteria predicting NIV failure would help clinicians choose between NIV and ETI as first-line treatment [1719]. In CPE-AOC patients NIV failure may independently prolong length of stay. Again, it is difficult to infer a causal relationship, but this finding may suggest a need for caution in order to avoid inappropriate delays in switching from NIV to ETI. In this population, however, the overall impact of NIV is clearly beneficial.

In conclusion, the use of NIV overall in patients with AOC or CPE and, when successful in all patients, is beneficial in terms of mortality, nosocomial pneumonia, and ICU length of stay. Nevertheless, NIV failure is associated with higher mortality in patients with de novo respiratory failure. This may at least indicate the need for some caution in this indication and for closely monitoring patients on NIV and switching promptly to ETI when necessary.