Study design, setting and ethical considerations

This is a multicenter retrospective cohort study between January, 2009 and December, 2012. There were 11 participating ICUs from a Brazilian network of private hospitals. One hospital is specialized in the care of heart diseases, while the others are mixed ICUs.

The Research and Ethics Committee of Hospital Pró-Cardíaco–the reference ethics committee designated by the National Research Ethics Committee–approved the retrospective analysis and publication of the data under the number 729.008 (CAAE: 33111214.3.0000.5533) and waived the need for informed consent given the retrospective nature of this study. We adhered to the STROBE guidelines (S1 File) and to the guidance from editors of respiratory, critical care and sleep journals [20].

Study population, exposures and outcome

The study population comprised very old patients (≥ 80 years old) admitted to the intensive care unit with CAP as the main reason for admission for whom invasive or noninvasive ventilation was deemed necessary. We excluded ICU readmissions. The exposure of interest was whether the initial respiratory support was NIV or IMV in the first day of ICU admission. The primary outcome was hospital mortality.

Data collection and definitions

We retrieved data from a prospectively collected multicenter ICU database (Epimed Monitor System®, Epimed Solutions®, Rio de Janeiro, Brazil), a cloud-based registry for ICU quality improvement in Brazil [21]. Data of all admitted patients are entered in the system by a trained case manager nurse and regularly audited. Retrieved variables included demographics (age and sex), body mass index, admission SAPS 3 [22], 1^st day sequential organ failure assessment (SOFA) score [23], previous functional status [7], comorbidities (all those from the Charlson comorbidity index) [24], 1^st hour physiological data (P_ao₂/F_io₂ ratio, P_aco₂ and pH), use of organ support in the 1^st hour, 1^st 24 hours and throughout ICU stay (vasopressors, renal replacement therapy, IMV, NIV), ICU and hospital length-of-stay and mortality, and palliative care decision within 24 hours of ICU admission. We calculated the Pneumonia severity index (PSI) [25], the modified frailty index (MFI) [26] and whether the patients were septic or not (according to Sepsis 3.0 definitions) [27, 28] from available variables of the dataset.

Data analysis

Categorical variables are described as numbers and percentages. Quantitative variables are presented as means (standard deviations) or medians [25^th percentile, 75^th percentile] accordingly. Continuous variables were evaluated for normality with histograms and analyzed with t-tests when appropriate. Non-normal continuous variables and discrete variables were analyzed with the Wilcoxon rank-sum test. Categorical variables were analyzed with the chi-squared test or the Fisher exact test, as appropriate.

The choice of the initial treatment strategy is associated with several clinical conditions. As recommended by Lederer et al., we chose the set of potential confounding factors based on a directed acyclic graph (DAG), accounting for causal paths and avoiding mediators, open back-door paths and collider-bias [20] (S1 Fig). We ran multivariable logistic regression models adjusting for potential confounders with different sets of covariates at a time: model 1: age and sex; model 2: model 1 plus body mass index (a proxy for malnutrition), COPD, heart failure, dementia and performance status; model 3: model 2 plus non-respiratory SOFA and admission source; model 4 (main model): model 3 plus respiratory variables (P_ao₂/F_io₂ < 150 mmHg, pH < 7.3, P_aco₂ > 50 mmHg).

We explored the robustness of the main model with four sensitivity analyses: 1) adding the SAPS 3 score in model 2, providing a different strategy to deal with severity and acuity variables spending less degrees of freedom; 2) excluding patients in whom palliative care was ascertained in the first day of ICU admission; 3) analyzing only patients admitted directly from the emergency department; and 4) running the main model (model 4) in complete case analysis. We also performed a post-hoc survival analysis censoring follow-up at 28 days of ICU admission. For this analysis, since we followed-up patients up to hospital discharge (or death), which would lead to informative censoring, we considered patients discharged alive as alive up to 28 days. We present results by their primary group (NIV vs. IMV) with a crude analysis (Kaplan-Meier survival curve and hazard ratios with 95% CIs from a univariate Cox model) and an adjusted analysis in a multivariable Cox model (adjusting for age, sex, modified frailty index and SAPS 3). We assessed proportional hazards by Schoenfeld residuals and log-log plots.

We evaluated effect modification by adding an interaction term in model 4 (main model) for ad-hoc defined subgroups: COPD, heart failure, severe functional status impairment, non-respiratory SOFA ≥ 4, P_ao₂/F_io₂ ≤ 150 mmHg and acute respiratory acidosis (pH < 7.3 and P_aco₂ ≥ 50 mmHg). Cut-off values were based on standard values used in other studies and traditional indications for IMV (eg, acute respiratory acidosis).

We derived the marginal prediction from model 4 to illustrate the association between the initial strategy of ventilation and mortality in some scenarios. The marginal prediction represents the mortality predicted by the model if all patients in the cohort had received NIV or IMV in each scenario, while all other covariates are kept as observed [29]. We explored scenarios that contrasted the magnitude of extrapulmonary organ dysfunction [low (0), intermediate (4) vs. high (8) non-respiratory SOFA score]; functional status impairment (severe vs. non-severe); and hypoxemia (P_ao₂/F_io₂ ratio > 150 or ≤ 150 mmHg). Furthermore, in an as-treated analysis, we evaluated the outcomes of patients of the NIV group who were intubated within or after 24 hours of ICU admission, and calculated adjusted odds ratios, adjusting for as model 4.

We conducted multiple imputations to deal with missing data in covariates [30]. We assumed data to be missing at random and imputed 50 datasets using chained equations with predictive mean matching. We included in the imputation model the outcome and exposure variables, all variables considered for covariate adjustment, interaction terms and auxiliary variables [31]. Further details of data missingness and the imputation model are described in the S1 Checklist.

We considered a p-value <0.05 as statistically significant for all analyses and no multiplicity adjustment was done. StataSE® version 16.0 was used for all analyses and the user-written mimrgns command was used to generate marginal effects.

Main findings

In this multicenter cohort study, NIV as the initial ventilatory strategy was not associated with lower hospital mortality among VOPs admitted to the ICU with CAP. We observed a significant degree of positive confounding, mainly explained by acuity variables included in the SOFA score. There was no strong evidence of effect modification regarding important subgroups of patients who usually benefit from NIV (COPD, heart failure) [12, 13] neither on those who usually have worse outcomes under NIV in this setting (P_ao₂/F_io₂ < 150, higher non-respiratory SOFA scores) [4]. Results were robust to sensitivity analyses of the main model and the imputation model assumptions. Irrespective of the initial ventilatory strategy, mortality rates were high for patients who were ultimately intubated, especially in the context of high non-respiratory SOFA scores, low P_ao₂/F_io₂ ratio and in patients who were intubated after the first 24 hours of ICU admission. We stress that these results should be interpreted as associations and should not imply causation, since residual confounding may still be an issue.

Relationship with the literature

The hospital mortality for VOPs admitted with CAP to the ICU is quite high. In our cohort, the overall hospital mortality was 55%. In the United States, the 30-day mortality of elderly (≥65 years-old) Medicare beneficiaries admitted with pneumonia to the ICU is of 36% [32], while for those who used NIV or IMV, 30-day mortality was 52.1% and 55.3%, respectively [33]. In the UK case-mix program, data from the early 2000’s show a hospital mortality of 50% regardless of age. In the study by Schortgen et al. [34], hospital mortality was 40% and 6-month mortality was 50% for VOPs undergoing NIV irrespective of the initial diagnosis. In this context, we primarily hypothesized that a less invasive strategy would benefit this high mortality target population, but our results were neutral.

One explanation is that the use of NIV in patients with high respiratory drive and high tidal volumes might be a matter of concern [35]. Recent evidence suggests that self-inflicted lung injury may occur through many mechanisms and could be an issue if one decides not to intubate and allow the patient to have substantial spontaneous respiratory efforts [36]. However, noninvasive ventilation is associated with reduced mortality in acute hypoxemic respiratory failure [37] and finding the patient who would be harmed by spontaneous breathing is still a matter of debate, though, including those under IMV [38].

An alternative explanation to our findings is that the strategy of mechanical ventilation may not have an impact on the outcome whatsoever if used within reasonable clinical grounds and considering that NIV is frequently used as ceiling therapy [34]. Valley et al. [33], in a large retrospective cohort of elderly (≥65 years) patients, showed that NIV was no better than IMV, with an absolute risk reduction in 30-day mortality of –0.7% (95% CI, -13.8–12.4, p = 0.92). Our results add to this literature in that we had more granular physiological data beyond ICD-9 codes and we could observe that acuity variables were the main confounders of this association and probably the main determinants of outcome. Although point-estimates did not differ much between NIV and IMV (Fig 3), there was an evident gradient of increased mortality from the confounding variables.

Another concern is that studies assessing NIV in patients with pneumonia have yielded varying results [3, 6, 33, 39], suggesting that NIV could even be harmful if applied to patients with a higher burden of non-respiratory organ dysfunctions or too hypoxemic [4, 5]. Our subgroup analyses assessing effect modification did not suggest this was the case. Point estimates were not worrisome, except for patients with a P_ao₂/F_io₂ ratio ≤ 150, in whom the point estimate differed substantially showing increased harm in those who received NIV, although with a high degree of uncertainty. Although this is biologically plausible and has been shown to be a predictor of failure of NIV, any interpretation of these findings should be very cautious, given the usual low power of subgroup analyses.

Our results should also be interpreted accounting for a risk of misdiagnosis of pneumonia [40]. Common explanations for misdiagnosis include both acute exacerbations of COPD or pulmonary edema from congestive heart failure, which may be difficult to differentiate from pneumonia at patient presentation. Although misdiagnosis could have a potential to bias our results, it would probably do so towards positive associations favoring NIV, which was not the case in our sample. Having said that, we believe this situation actually reflects usual clinical practice and may enhance the generalizability of our findings [41].

Implications for practice

Our results suggest that a careful balance of benefits and harms of each strategy and the risks of NIV failure–and worse outcomes when in happens–should be weighed against patient’s values and preferences to decide the best course of action. Clinicians should not avoid using NIV as an initial strategy of ventilation if it is used within reasonable clinical grounds, since there was no strong signal of harm from this strategy. Furthermore, NIV use as ceiling therapy should still be regarded as an alternative to IMV when treatment limitations are in place. By contrast, clinicians should not avoid offering IMV for VOPs admitted with pneumonia, especially early during their ICU course. Nevertheless, the prognostic implications of the need for IMV should be considered, especially later during ICU stay and when the burden of extra-pulmonary organ dysfunction is higher and functional impairment is severe.

Strengths and limitations

Our manuscript has some strengths. The multicenter nature of our study with 11 ICUs from different hospitals and the real-world scenario with treatment limitations in place enhances the generalizability of our results. To account for the inherent risks of an observational study, we developed a causal DAG, which is an important step towards stronger inference and explicit selection of confounders [20], with sensitivity analyses to assumptions of the model being robust. Furthermore, we used multiple imputation to account for data missingness, a recommended approach that usually leads to less biased estimates while retaining higher statistical power [30, 31].

These results are also amenable to limitations. Although we collected data from 11 ICUs over four years, we might have had low power to detect an association of benefit for NIV over IMV, considering we observed a protective point-estimate on our adjusted models; however, we had enough data points to include the main confounders in the regression model without overfitting the data. Our data dates back to 2009–2012, which could be an issue, but except for prone positioning in severe acute respiratory distress syndrome [42], no other therapies appliable to very elderly patients [43] received a strong recommendation since then. We did not have data on pneumonia etiology and specific treatment, which might have an impact on outcome; nevertheless, we do not believe this would explicitly influence clinician’s decision to use NIV or IMV and would therefore not be in the DAG. We also did not have access to how NIV or IMV were specifically deployed, such as NIV mode and interfaces and how PEEP was titrated. Furthermore, there could be variability between hospitals because of expertise with NIV, which was not specifically addressed in our models. Hospital mortality may not be the best outcome to be assessed in this population, since long-term functional outcomes could be more valuable; unfortunately, we could not have access to such outcomes. Data on do-not-intubate orders were not explicitly available, although these are seldom used in Brazil and this could be partially captured when excluding patients under palliative care. We also did not have measures of arterial blood gases neither ventilatory variables after the start of NIV, which could provide more mechanistic insights to explain our results [5, 35].