Authors: Fredy Leonardo Carreño-Hernández, Sergio Prieto, Daniela Abondando, Jairo Alejandro Gaitán, Yenny Rocío Cárdenas -Bolívar, Adriana Beltrán, Jorge Iván Alvarado-Sánchez, Joseph L. Nates
Categories: Research, Critical care, Mechanical ventilation, Noninvasive ventilation, COVID-19, SARS-CoV-2
Source: Systematic Reviews
Authors: Fredy Leonardo Carreño-Hernández, Sergio Prieto, Daniela Abondando, Jairo Alejandro Gaitán, Yenny Rocío Cárdenas -Bolívar, Adriana Beltrán, Jorge Iván Alvarado-Sánchez, Joseph L. Nates
The COVID-19 pandemic has resulted in a critical shortage of respiratory ventilators, highlighting the urgent need to explore alternative treatment options for patients with acute respiratory distress syndrome (ARDS) caused by respiratory viruses, as an alternative to invasive mechanical ventilation (IMV) in future pandemics.
The objective of this study was to assess the effectiveness of alternative noninvasive oxygenation and ventilation strategies in comparison to invasive mechanical ventilation (IMV) in patients with virus-induced acute respiratory failure (ARF). The primary outcome was the all-cause ICU mortality rate.
A systematic review was conducted following the Cochrane guidelines and PRISMA reporting guidelines. The search encompassed databases such as Medline, Cochrane CENTRAL, and Embase to identify relevant indexed literature. Additionally, gray literature was included by consulting regulatory agencies. The included studies compared various oxygenation and ventilatory alternatives, such as high-flow nasal cannula (HFNC), continuous positive airway pressure (CPAP), or noninvasive mechanical ventilation (NIMV) with IMV. An exploratory meta-analysis was performed by calculating the risk ratio (RR) by random effects and meta-regression to explore possible sources of heterogeneity and to compare ventilatory alternatives against IMV to reduce mortality, length of stay (LOS) days in ICU, nosocomial infection, and barotrauma.
A total of forty-seven studies were included in this systematic review. NIMV had an RR of 0.70 (0.58–0.85), HFNC had an RR of 0.54 (0.42–0.71), and CPAP had an RR of 0.80 (0.71–0.90), with meta-regression models that reduced heterogeneity to 0%. For LOS days in ICU, NIMV had 0.38 (− 0.69: − 0.08) lower days and HFNC 0.29 (− 0.64: 0.06) lower days with meta-regression models that reduction heterogeneity to 0% for HFNC and 50% for NIMV. Not enough studies reported nosocomial infection or barotrauma to evaluate them in a meta-analysis. The overall quality of evidence, as assessed by GRADE evaluation, was determined to be from very low to medium certainty depending on the ventilatory strategy and outcome.
The findings of this systematic review support the use of alternative noninvasive oxygenation and ventilation strategies as viable alternatives to conventional respiratory ventilation for managing viral-induced ARF. Although it is essential to interpret these findings with caution given the overall low to medium certainty of the evidence, the integration of these modalities as part of the management strategies of these patients could help reduce the utilization of ICU beds, invasive ventilators, and costs in both developed and developing countries.
The online version contains supplementary material available at 10.1186/s13643-025-02775-6.
The COVID-19 pandemic has brought forth unprecedented challenges in the management of critically ill patients, especially within the confines of intensive care units (ICUs), where the provision of ventilatory support holds pivotal importance [1]. Among these patients, acute respiratory failure (ARF) resulting from COVID-19 and other viral infections is the leading cause of mortality [2]. Additionally, the global shortage of respiratory ventilators, as underscored by the World Health Organization (WHO), has further intensified the severity of the situation [2]. The elevated mortality rates observed among ICU patients requiring ventilatory support emphasize the pressing need for effective interventions [3].
Numerous ventilatory strategies have been described to address the medical management of ARF, including noninvasive ventilation (NIV), invasive mechanical ventilation (IMV), and standard oxygen therapy (SOT). Notably, compared with SOT and IMV, NIV has been shown to decrease mortality in patients with acute respiratory distress syndrome (ARDS), a form of ARF [4]. Furthermore, the utilization of NIV in COVID-19 patients has shown promise in reducing mortality [5]. A systematic review provided evidence supporting the efficacy of NIV over IMV in managing acute hypoxic respiratory failure caused by coronaviruses, such as COVID-19, severe acute respiratory syndrome (SARS), and Middle East respiratory syndrome (MERS). However, that review was limited in scope, as it did not encompass patients with ARDS caused by H1N1 [5].
To address these gaps and contribute to the understanding of optimal oxygenation and ventilation strategies, this study aimed to comprehensively evaluate the efficacy and safety of alternative techniques. These included HFNC therapy, CPAP therapy, NIV, ventilator splitters (VS), and low-cost ventilators (LCVs) for the management of ARF caused by SARS-CoV-2, SARS, MERS, or H1N1. The primary outcome measure was all-cause ICU mortality, with a specific focus on mortality in cases of NIV failure, serving as a crucial safety indicator.
We conducted a systematic review adhering to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) framework [6] as a reporting guideline and followed the guidelines outlined in the Cochrane Handbook for Systematic Reviews of Interventions [7]. The study protocol was registered at PROSPERO (CRD42020199175).
Randomized clinical trials (RCTs), nonrandomized studies (NRSs) (cohorts and case‒controls), or congress abstracts, if they had results, were included. The inclusion criterion for identifying both published and unpublished studies was abstracts from congresses, which were indexed by the authors’ names. The study followed the PICO framework, which is structured as
Patients (P): This study concentrated on critically ill adult patients diagnosed with ARF attributed to SARS-CoV-2, SARS, MERS, or H1N1 infections. The inclusion criterion was patients who were monitored until ICU discharge, irrespective of their final outcome. Virus identification relies on diverse methods, including polymerase chain reaction (PCR), antigen testing, or epidemiological association. ARDS was defined in accordance with the Berlin criteria.
Intervention (I): The intervention included an array of respiratory support methods, including high flow nasal cannula (HFNC), continuous positive airway pressure (CPAP), NIMV, and invasive strategies such as ventilator splitters or LCV. The following specific definitions were established for each HFNC: Continuous flow at or exceeding 10 ml/min, alongside specified temperature and humidification parameters.CPAP: Continuous pressure equal to or exceeding 5 mmHg delivered via a facemask.LCV: Mechanical compression of resuscitation bags.Ventilator splitters (VS): Devices with or without resistance valves.NIMV: Ventilation using masks or helmets, encompassing noninvasive positive-pressure ventilation or bilevel positive airway pressure (BiPAP). These precise definitions ensured consistency and clarity in describing the distinct respiratory support methods utilized in the study.
Comparator (C): The comparator used was invasive mechanical ventilation (IMV).
Outcome (O): The primary outcome of interest was all-cause ICU mortality. Secondary outcomes where length of stay in ICU (LOS), barotrauma, and nosocomial infection.
The researchers excluded patients from registers that involved ARF caused by other factors, cases where the viral infection was not confirmed, case series, case reports, consensus papers, or articles written in languages other than Spanish or English.
The comprehensive search strategy for this systematic review can be found in the PROSPERO protocol with the registration code CRD42020199175. The EMBASE, MEDLINE, and Cochrane CENTRAL databases were searched between January 2000 and August 2021 and between September 2021 and December 2022. The retrieved articles were downloaded and subjected to manual selection using the Rayyan™ platform to identify relevant indexed literature.
To include gray literature, additional research was performed on regulatory and governmental agency websites such as clinicaltrial.gov, the Food and Drug Administration (FDA), and the “Instituto Nacional de Vigilancia de Medicamentos y Alimentos (INVIMA)”. For studies on trials, manual selection was conducted based on the engineered searches, as the downloadable content was not compatible with the Rayyan™ platform.
In the research process, three authors (C.F.L., G.J.A., and A.D.) reviewed the abstracts based on the predefined inclusion criteria. Afterward, two authors (C.F.L. and P.S.) independently reviewed the full texts of the selected articles. Studies that met the inclusion criteria were included in the data extraction process. In case of any discrepancies or disagreements, the authors resolved them through consensus among themselves.
A standardized data table form in Microsoft Excel was developed to include all the data, and two authors extracted the information (C.F.L.; A.D.). The extracted data included country, study design, year, virus, intervention at admission, total patients, patients in the comparator, patients on intervention with subsequent intubation, ventilatory parameters, PaO2/FiO2, rate oxygenation (ROX) index, APACHE II score, smoking habit, hypertension, ARDS classification, age, antiviral and anti-inflammatory treatment, and outcomes such as mortality in the ICU, LOS ICU, barotrauma, and nosocomial infection. Continuous variables are reported as medians or medians with their respective standard deviations or interquartile ranges, respectively, while categorical variables are reported as frequencies and percentages. Incomplete data were registered as “no data” or extrapolated from the frequency of patients in interventions and IMV if possible. For studies in trials, the extracted data included the study reference, country, study design, intervention, and phase.
Two independent authors (C.F.L.; P.S.) assessed the risk of bias using the Cochrane Risk of Bias (RoB) tool for randomized controlled trials (RCTs) or the Risk of Bias in Nonrandomized Studies of Interventions (ROBINS-I) tool for nonrandomized studies (NRS) [8–10]. For RoB, it was evaluated the random sequence generation, the allocation concealment, the binding of participants and personnel, the blinding of outcomes, incomplete outcome data, and selective reporting. For ROBINS-I bias for confounding, in selection of participants in the study, in classification of interventions, in deviations from intended interventions, due to missing data, in measurement of outcomes and in selection of the reported results were evaluated. No bias assessment was conducted for abstracts presented at congresses or trials on course due to incomplete information.
Meta-analysis was conducted using a random effects model for each ventilatory strategy. For dichotomous outcome RR was calculated and mean differences for continues outcomes were obtained, both with confidence intervals 95% (CIs) calculated. One analysis was performed combining all RcT and NRS, while other analysis only between same studies design was performed. Heterogeneity was evaluated by the Cochran Q statistic, and its effect was measured using inconsistency [I^2^] [10]. Sensitivity analysis was performed by excluding studies with high risk of bias to evaluate the impact of bias in results. Asymmetry was visually evaluated by funnel plots and by Egger’s test for each intervention with and without abstract congress for publication bias evaluation. The data were analyzed using R version 4.2.2 with meta-packages, and significance level of P < 0.05 was considered to indicate statistical significance.
The meta-regression analysis included age, PaO2/FiO2, virus type, SOFA score, and APACHE II score means for interventions. Studies reporting medians with interquartile ranges or maximum and minimum values were transformed into means with standard deviations using Sean McGrath’s Box‒Cox method. Incomplete data were imputed out (generated) using the "mice" package and the "cart" method so missing data can be filled. The sensitivity analysis involved various combinations of included or excluded abstracts, with or without imputed data and with only those studies that had the lowest risk of bias. The chosen models were based on covariate combinations with the lowest heterogeneity.
To evaluate the quality of evidence and formulate clinical recommendations, the researchers employed the Grading of Recommendations, Assessment, Development, and Evaluations (GRADE) approach [11, 12]. The components evaluated in GRADE were the study design, risk of bias, inconsistency, indirectness, imprecision, and other considerations. To facilitate the grading process and create evidence profiles, the researchers utilized GRADEpro™ software.
A total of 5739 records were initially identified through the literature search. After applying the inclusion and exclusion criteria, 47 studies were included in the review (Fig. 1). Of these, 36 were nonrandomized studies (NRS) [13–49], 2 were trial studies [50, 51], and 9 were abstracts from congresses [52–60]. As of December 2023, no published articles derived from the Abstract congresses included were identified.Fig. 1PRISMA flowgram. FDA: Food and Drug Administration. INVIMA: Instituto Nacional de Vigilancia de Medicamentos y Alimentos
The study included 36 NRS and nine abstract congress studies. Among these, 12 studies concentrated on NIMV [15, 18, 20, 21, 23, 26–29, 40, 46, 60], seven on CPAP [19, 33, 41, 44, 55, 57, 59], and six on HFNC [25, 32, 36, 39, 47, 53]. Five studies evaluated multiple interventions, four involving NIMV and HFNC [13, 14, 30, 60] and one involving NIMV and CPAP [34]. Fifteen studies did not distinctly differentiate between interventions. Among the 36 NRS studies, 16 were prospective cohort studies [13–15, 17–21, 23–28, 41, 49], 19 were retrospective cohort studies [22, 29–46], and one was a nested case‒control study [48]. Various clinical characteristics were reported across the studies, including ventilatory parameters, age, frequency of hypertension, ICU days, SOFA score, APACHE II score, PaO2/FiO2 levels, anti-inflammatory and antiviral treatments, smoking habits, ARDS classification, barotrauma, and nosocomial superinfection. Additionally, ongoing trials related to noninvasive ventilation (NIV) and spontaneous ventilation (SV) have been conducted with no results published. Detailed information on each study’s characteristics and evaluated variables can be found in Table 1 and the online supplement (Table S1).
Table 1Studies general characteristicsAuthorDesign CountyVirusInterventions PopulationVentilatory parametersAPACHE-II ScoreSOFA scoreAGEPaO2/FiO2RR [IC95%]1. Asghar 2020 [17]Prospective cohortPakistanSARS-CoV-2IMV: 50CPAP: 59No dataNo dataNo dataNo dataNo data-0.60 [0.45, 0.80]2. Bertaina 2020 [18]Prospective cohortMulticountrySARS-CoV-2IMV: 106NIMV: 390No dataNo dataNo data67 (57–73)70 (58–79)No data-0.49 [0.42, 0.58]3. Brink 2012 [19]Prospective cohortSwedenH1N1IMV: 26CPAP: 67FiO2: 90%PEEP: 5TDV: 8.9No dataNo data52 (38–58)49 (38–59)80 (40–108)86 (63–116)-2.20 [0.70, 6.88]4. Garcia 2021 [13]Prospective cohortSwitzerlandSARS-CoV-2IMV: 92NIMV: 87NIMV-F: 43HFNC: 87HFNC-F: 45FiO2: 63%PEEP: 12TDV: 6.1FiO2: 41%PEEP: 12TDV: 6.1FiO2: 60%11 (8–20)10 (7–16)No data10 (6–13)No data7 (6–8)6 (4–7)No data6 (3–7)No data62 (55–70)66 (55–76)No data63 (55–74)No data123 (90–165)135 (97–168)No data126 (79–169)No data-1.47 [0.94, 2.30]1.49 [0.88, 2.52]0.78 [0.45, 1.36]1.24 [0.71, 2.18]5. Masclans 2013 [20]Prospective cohortSpainH1N1IMV: 312NIMV: 177NIMV-F: 105No data15 (10,–19)12 (9–16)14 (10–18.2)No data43 (32–52)44 (33–53)44 (32–53)No data-0.66 [0.44, 0.97]1.11 [0.76, 1.61]6. Mellado-Artigas 2021 [49]Prospective cohortSpainSARS-CoV-2IMV (A): 61IMV (B): 251HFNC (A): 61HFNC-F (A): 21HFNC (B): 95FiO2: 79%O2 L/ 55FiO2: 72%11 (9–14)12 (10–14)10 (9–113)No data10 (7–13)5 [3,–7]7 [5–8]4 [4–7]No data4 [3–5]61 (11)62.3 (10.9)62 (11)No data60.6 (12.7)117 (51)121.2 (58.0)121 (49)No data122.7 (48.5)--0.69 [0.32, 1.50]1.34 [0.58, 3.08]0.34 [0.19, 0.60]7. Mukhtar 2020 [21]Prospective cohortEgyptSARS-CoV-2IMV: 9NIMV: 30No data10 ± 411 ± 5.5No dataNo data175 (118–205)170 (112–224)-0.13 [0.04, 0.40]8. Polok 2022 [23]Prospective cohortMulticountrySARS-CoV-2IMV: 1876NIMV: 630No dataNo data7 (4–8)4 (3–6)75.1 ± 4.276.8 ± 4.9No data-1.01 [0.92, 1.11]9. Rama-Maceiras 2021 [24]Prospective cohortSpainSARS-CoV-2IMV: 123NIV: 214NIV-F: 110No dataNo dataNo dataNo dataNo data-0.46 [0.29, 0.74]0.90 [0.57, 1.43]10. Rello 2012 [25]Prospective cohortSpainH1N1IMV: 10HFNC: 20HFNC-F: 10O2 L/ 25O2 L/ 3025 (13–31)11 (8–14)No data8 (4–12)3 (3–5)No data50 (44–56)37 (29–47)No data73 (56–81)135 (84–210)No data-0.75 [0.15, 3.79]1.50 [0.32, 7.14]11. Reyes 2022 [14]Prospective cohortMulticountrySARS-CoV-2IMV: 14,197NIMV: 24,112NIMV-F: 6749HFNC: 28,256HFNC-F: 3538No dataNo dataNo dataNo dataNo data-1.03 [1.00, 1.05]1.30 [1.26, 1.35]0.91 [0.89, 0.94]1.13 [1.09, 1.19]12. Ríos 2009 [26]Prospective cohortArgentinaH1N1IMV: 129NIMV: 49FiO2: 80% ^a^PEEP: 10 ^a^TDV: 7.8 ^a^PP: 26 ^a^19 ± 716 ± 6No data44 ± 1540 ± 16129 (70 to 173)141 (92 to 217)-0.76 [0.51, 1.12]13. Rodríguez 2017 [15]Prospective cohortSpainInfluenzaIMV: 1092NIMV: 806NIMV-F: 458No data7 (5–10)7 (4–9)4 (3–6)7 (5–10)7 (4–9)4 (3–6)51 (39–62)58 (47–68)53 (42–65)No data-0.78 [0.68, 0.91]1.23 [1.06, 1.42]14. Sivaloganathan 2020 [27]Prospective cohortUKSARS-CoV-2IMV: 21NIMV: 82NIMV-F: 27No data22 (15–25)11 (8–12.5)18 (13.0–24.5)6 (4–8)3 (4–3)4 (3–6)61 (18–65)50 (45–60)57 (50–64)No data-0.98 [0.46, 2.10]0.39 [0.11, 1.38]15. Yam 2005 [28]Prospective cohortChinaSARS-CoV1IMV: 451NIMV: 42No dataNo dataNo data4447No data-0.29 [0.09, 0.86]16. Alraddadi 2019 [29]Retrospective cohortSaudi ArabianMERSIMV: 197NIMV: 105NIMV-F: 97No dataNo data7 (4–9)9 (7–12)7 (4–9)58 (45–69)60 (50–73)61.0 (52–73)106 (68, 166)110 (62, 160)103 (62, 160)-0.88 [0.75, 1.02]0.95 [0.82, 1.10]17. Baqi 2021 [30]Retrospective cohortPakistanSARS-CoV-2IMV: 57NIMV: 100HFNC: 21No dataNo dataNo data60 (50–65)No data-0.77 [0.67, 0.89]0.51 [0.33, 0.81]18. Berenguer 2020 [31]Retrospective cohortSpainSARS-CoV-2IMV: 619NIV: 528No dataNo dataNo data70 ^a^No data-1.00 [0.88, 1.14]19. Castro 2022 [32]Retrospective cohortSpainSARS-CoV-2IMV: 101HFNC: 84HFNC-F: 52No dataNo dataNo data127 ± 55148 ± 64No data64 ± 1466 ± 12No data-0.55 [0.36, 0.84]0.90 [0.49, 1.63]20. Domenico 2021 [33]Retrospective cohortItalySARS-CoV-2IMV: 13CPAP: 90CPAP-F: 36No dataNo dataNo data56.231 ± 3.4458.444 ± 2.79No data194.09 ± 29.81247.87 ± 16.73202.82 ± 16.88-0.69 [0.49, 0.99]0.61 [0.39, 0.97]21. Duca 2020 [34]Retrospective cohortItalySARS-CoV-2IMV: 7NIMV: 7CPAP: 71No dataNo dataNo data64 (62–72)72(59–80)70 (62–79)76 (60–177)87 (53–120)131 (97—190)-0.60 [0.33, 1.08]0.76 [0.67, 0.87]22. Forrest 2021 [35]Retrospective cohortGermanySARS-CoV-2IMV:154NIV: 534No dataNo dataNo data68 ± 1367 ± 16No data-0.38 [0.33, 0.44]23. Hernandez 2020 [36]Retrospective cohortUSASARS-CoV-2IMV: 97HFNC: 109HFNC-F: 78No dataNo dataNo data7 (5–10)9 (7–11)No data62 (55–73)65 (56–75)148 (111–205) ^a^-0.55 [0.36, 0.84]0.77 [0.51, 1.16]24. Hesselle 2022 [37]Retrospective cohortGermanySARS-CoV-2IMV: 492NIV: 87No dataNo dataNo dataNo dataNo data-0.95 [0.71, 1.26]25. Hua 2020 [38]Retrospective cohortChinaSARS-CoV-2IMV: 113NIV: 152No dataNo data6.0 ± 3.05.5 ± 2.767 ± 1371 ± 10No data-0.44 [0.36, 0.54]26. Lee 2020 [39]Retrospective cohortSouth KoreaSARS-CoV-2IMV: 23HFNC: 24HFNC-F: 16No data15 (10–17)11 (8–14)14 (8–15)3 (2–7)2 (2–4)3 (2–4)72 (64–76)69 (60–78)66 (59–77)86 (69–123)144 (70–206)120 (62–188)-0.59 [0.30, 1.15]0.77 [0.40, 1.50]27. Jamil 2021 [40]Retrospective cohortPakistanSARS-CoV-2IMV: 33CPAP: 28No dataNo dataNo data57.06 ± 12.6353.75 ± 18.10No data-0.69 [0.45, 1.05]28. Myers 2023 [22]Retrospective cohortUSASARS-CoV-2IMV: 2965NIV: 2793No dataNo dataNo data60.8 ± 15.362.4 ± 12.7No data-0.32 [0.30, 0.35]29. Pasin 2022 [41]Retrospective cohortUSASARS-CoV-2IMV: 141CPAP: 141CPAP-F: 89No dataNo data4 (3–5)4 (3–5)68 (58–75)66 (57–73)114 (74–195)120 (83–181)No data-0.61 [0.42, 0.89]0.81 [0.55, 1.19]30. Pia 2021 [42]Retrospective cohortUSASARS-CoV-2IMV: 91NIV: 131NIV-F: 34No dataNo dataNo data67 (60–76)67 (65–82)69 (58–75)No data-0.90 [0.78, 1.03]1.02 [0.87, 1.20]31. Rizwan 2022 [43]Retrospective cohortPakistanSARS-CoV-2IMV: 30NIV: 44No dataNo dataNo dataNo dataNo data-0.68 [0.53, 0.87]32. Rocans 2022 [44]Retrospective cohortLatviaSARS-CoV-2IMV: 156CPAP: 374CPAP-F: 235No dataNo dataNo data63 (62–63)65 (63–66)63 (61.0–65.4)No data-0.91 [0.47, 1.09]3.23 [0.81, 1.02]33. Rajdev 2021 [45]Retrospective cohortUSASARS-CoV-2IMV: 121NIV: 232TDV: 6.4 (0.5)FiO2: 93%PEEP: 9.28No dataNo dataNo dataNo data-0.15 [0.05, 0.44]34. Wang 2020 [46]Retrospective cohortChinaSARS-CoV-2IMV: 50NIMV: 91No dataNo dataNo data64 (55–70) ^a^108 (58–219)261 (175–287)-0.42 [0.26, 0.68]35. Zheng 2020 [47]Retrospective cohortChinaSARS-CoV-2IMV: 15HFNC: 19No dataNo dataNo data71 (60, 83)66 (51, 72)No dataNotestimable36. Hamouri 2021 [48]Case‒ControlJordanSARS-CoV-2IMV: 112CPAP: 127TDV: 6.18PP: 29PEEP: 16No dataNo data67.7 ± 14.464.7 ± 12.8No data-Not estimable37. Woo 2020 [50]On trial – withdrawUSASARS-CoV-2IMV: no dataSV: no dataNo dataNo dataNo dataNo dataNo dataNo data38. Keith 2020 [51]On trial –not recruitingUSASARS-CoV-2IMV: no dataNIV: no dataNo dataNo dataNo dataNo dataNo dataNo data39. Buddharaju 2021 [52]Abstract congressUSASARS-CoV-2IMV: 35NIMV: 23No dataNo dataNo dataNo dataNo data-0.38 [0.14, 0.99]40. Chowdhury 2021 [53]Abstract congressUSASARS-CoV-2IMV: 49HFNC: 125No dataNo dataNo dataNo dataNo data-0.55 [0.30, 0.84]41. Chang 2022 [54]Abstract congressUSASARS-CoV-2IMV: 235NIV: 398No dataNo dataNo dataNo dataNo data-0.65 [0.51, 0.83]42. Eswarappa 2021 [55]Abstract congressUKSARS-CoV-2IMV: 80CPAP: 215CPAP-F: 31No dataNo dataNo dataNo dataNo data-0.72 [0.54, 0.97]1.36 [0.96, 1.92]43. Flaatten 2021 [56]Abstract congressNorwaySARS-CoV-2IMV: 1777NIV: 690NIV-F: 331No data18.77.2176.84.35.675.477.375.6No data-1.66 [1.55, 1.78]0.77 [0.65, 0.91]44. Iftikhar 2021 [57]Abstract congressUKSARS-CoV-2IMV: 34CPAP: 95CPAP-F: 24No dataNo dataNo data58.8 ± 13.261.7 ± 11.556 ± 12.8No data-0.86 [0.49, 1.39]1.98 [0.50, 1.92]45. Mourisco 2022 [58]Abstract congressPortugalSARS-CoV-2IMV: 47NIV: 102No dataNo dataNo data70 ± 1565 ± 1599 ± 52104 ± 66-0.89 [0.50, 1.57]46. Napolitani 2021 [59]Abstract congressItalySARS-CoV-2IMV: 44CPAP: 22No dataNo dataNo dataNo dataNo data-0.85 [0.52, 1.37]47. Yamamoto 2022 [60]Abstract congressJapanSARS-CoV-2IMV: 1139NIMV: 248HFNC: 1072No dataNo dataNo dataNo dataNo data-0.44 [0.30, 0.64]0.28 [0.22, 0.36]IMV invasive mechanical ventilation, NIV non-invasive ventilation, NIMV non-invasive mechanical ventilation, HFNC high flow nasal cannula, CPAP continuous positive airway pressure, SV splitter ventilator, NIV-F non-invasive ventilation failure, NIMV-F non-invasive mechanical ventilation failure, HFNC-F high flow nasal cannula failure, CPAP-F continuous positive airway pressure failure, PEEP positive end expiratory pressure (cm H2O), PP plateau pressure (cmH2O), FiO2 fraction of inspired oxygen, TDV tidal volume (ml/kg)^a^No mean or median per group but general
The analysis examined 36 NRS studies using the ROBINS-I since RCTs were absent. No bias assessment was performed for nine congress abstracts or two trial studies. Among the 16 NIMV studies, most had a low risk of bias due to missing data, in the measurement of outcomes and in the selection of the reported result, but only some overlooked critical risk factors such as age and disease severity. Nine HFNC studies were evaluated; although all had a low risk of bias, such as NIMV, some did not adjust for important risk factors. CPAP studies varied, with only a few having a low risk of bias in all domains. Additionally, eight studies lacking clear intervention distinctions were assessed. The details are shown in Fig. 2 and individual details for risk of bias assessments are reported in table S2.Fig. 2ROBINS-I evaluation for all ventilatory strategies
As no RCT was found, only NRS were considered for syntheses. In the comparison between NIMV and IMV, 15 studies and one abstract study revealed a statistically significant reduction in mortality, with an RR of 0.70 [95% CI 0.58–0.85; I^2^ = 91%] (Fig. 3A). By only considering 8 studies with the lowest bias for the sensitivity analysis, similar results were obtained with an RR of 0.72 [95% CI 0.52–0.99; I^2^ = 91%]. The same results were obtained when abstracts were excluded, with an RR of 0.73 [95% CI 0.60–0.80; I^2^ = 90%] (Fig. 3B). When assessing cases where NIMV failed and required subsequent intubation versus IMV, the analysis of seven studies revealed no significant differences in mortality (RR 1.07 [95% CI 0.89–1.29; I^2^ = 85%]) (Fig. 3C). Four studies with the lowest bias were considered for sensitivity analysis having similar results, with an RR of 0.97 [95% CI 0.78–1.21; I^2^ = 56%].
Fig. 3A Forest plot of the difference between NIMV and IMV for mortality with abstract congress. B Forest plot of the difference between NIMV and IMV for mortality without abstract congress. C Forest plot of the relationship between NIMV failure and IMV mortality
For HFNC therapy compared to IMV, a meta-analysis of eight studies and two abstract congresses demonstrated a substantial decrease in mortality (RR 0.54 [95% CI 0.42–0.71; I^2^ = 91%]) (Fig. 4A). No differences were detected when abstracts were excluded (RR 0.63 [95% CI 0.49–0.80; I^2^ = 71%]) (Fig. 4B). Similar results for sensitivity analysis were obtained with 3 studies with the lowest bias risk (RR 0.65 [95% CI 0.46–0.92; I^2^ = 0%]). Conversely, when HFNC therapy failed and led to intubation versus IMV, seven studies revealed no statistically significant differences in mortality (OR 1.05 [95% CI 0.88–1.25; I^2^ = 0%]) (Fig. 4C). No differences were detected in the sensitivity analysis of mortality among the 3 studies with the lowest risk of bias (RR 0.98 [95% CI 0.69–1.38; I^2^ = 0%]).Fig. 4A Forest plot of HFNC therapy against IMV for mortality with an abstract congress. B Forest plot of HFNC therapy against IMV for mortality without an abstract congress. C Forest plot of HFNC failure versus IMV for mortality
Regarding CPAP, a meta-analysis of five studies and three abstracts revealed a reduction in mortality (RR 0.80 [95% CI 0.71–0.90; I^2^ = 37%]) (Fig. 5A). No changes were identified when abstracts were excluded (RR 0.80 [95% CI 0.68–0.93; I^2^ = 62%]) (Fig. 5B). The same results were obtained using only two studies with the lowest risk of bias for the sensitivity analysis (RR 0.73 [95% CI: 0.62–0.87; I^2^ = 18%]). In cases where CPAP failed and resulted in intubation versus IMV, the analysis of seven studies and two abstract congresses demonstrated an increase in mortality, although the confidence interval was wider (RR 0.94 [95% CI 0.70–1.27; I^2^ = 80%]) (Fig. 5C), with no differences if abstract congress was excluded (RR 0.86 [95% CI: 0.56–1.26; I^2^ = 86%]) (Fig. 5D). Important differences were identified for the sensitivity analysis using two studies with the lowest risk of bias (RR 0.67 [95% CI 0.49–0.92; I^2^ = 36%]).Fig. 5A Forest plot of the association between CPAP and IMV in terms of mortality with abstract congress. B Forest plot of the association between CPAP and IMV for mortality without an abstract congress. C Forest plot of mortality in patients who experienced CPAP failure versus those who experienced IMV with abstract congress. D Forest plot of mortality without abstract congress in patients who experienced CPAP failure versus those who experienced IMV
Calculated unadjusted RR for each study used in the meta-analysis can be found in Figs. 3A–C, 4A–C, and 5A–D, while adjusted RR reported by the corresponding authors in the included studies can be found in table S1.
In the comparison between NIMV and IMV, 6 NRS and no abstract congress revealed a statistically significant reduction in LOS mean, with a difference of − 0.38 days [95% CI − 0.69: − 0.08; I^2^ = 76%] (Fig. 6A). By considering only 5 studies with the lowest risk of bias for sensitivity analysis, the LOS mean difference was the same with − 0.27 days [95% CI − 0.52: − 0.06; I^2^ = 71%]. In the case of HFNC against IMV with 4 NRS and no abstract congress, no significant differences were identified as mean difference was − 0.29 days [95% CI − 0.64: 0.06; I^2^ = 81%] (Fig. 6B). Different results were obtained when only including studies with the lowest bias risk with a mean difference of 0.98 days [95% CI 0.69: 1.38; I^2^ = 0%] higher. No meta-analysis for CPAP was possible as no study report LOS days in ICU, as same goes to NIMV-failure and HFNC-failure as no study report LOS days exclusive for those scenarios. In the same way, not enough studies with the same ventilation strategy reported nosocomial infection or barotrauma to evaluate them in a meta-analysis. In the case of nosocomial infection, Masclans [20] report 60 (19.2%) cases in IMV and 20 (11.1%) cases for NIMV, while Wang [46] reported 14 (28.0%) cases for IMV and 15 (16.5%) for NIMV. For barotrauma, Brink [19] reported 4 (15%) cases in IMV and 4 (6%) cases in CPAP, while Hamouri [48] reported 17 (33.3%) cases in IMV and 34 (66.7%) cases in NIV. Fig. 6Forest plot of LOS days in ICU. A Forest plot of the association between NIMV and IMV in terms of LOS days in ICU. B Forest plot of the association between HFNC and IMV for LOS days in ICU
For the studies and abstracts that involved NIMV or HFNC therapy in mortality, asymmetry and risk of bias were identified against the intervention in their funnel plots (Fig. 7A, B). Different results were obtained for studies that involved CPAP, as asymmetry was not observed in their funnel plot (Fig. 7C). Similar results were obtained via Egger’s test, with NIMV showing a β0 coefficient of − 2.56 and HFNC exhibiting a β0 coefficient of − 2.54. For CPAP, Egger’s test was not possible because fewer than 10 studies were included. No differences were detected when abstracts were excluded (Fig. 7D–F). In LOS days in ICU, no significant asymmetry was observed in the funnel plots for NIMV and HFNC (Fig. 8A, B), but no Egger’s test was possible as the number of studies was too small to test for small study effects.Fig. 7A Funnel plot of the difference between NIMV and IMV for mortality with abstract congress. B Funnel plot of the difference between NIMV and IMV for mortality without abstract congress. C Funnel plot of HFNC therapy against IMV for mortality with an abstract congress. D Funnel plot of HFNC therapy against IMV for mortalitywithout an abstract congress. E Funnel plot of the association between CPAP and IMV mortality with abstract congress. F Funnel plot of the association between CPAP and IMV for mortality without an abstract congressFig. 8A Funnel plot of the difference between NIMV and IMV for LOS days in ICU. B Funnel plot of the difference between NIMV and HFNC for OS days in ICU
In the meta-regression analyses comparing different noninvasive ventilation (NIV) modalities with invasive mechanical ventilation (IMV) for mortality, specific factors were considered to reduce heterogeneity. For NIMV versus IMV, models that included age, PaO2/FiO2, and the specific virus type significantly reduced heterogeneity. The model excluding imputed data and abstracts proved most effective. In the case of NIMV failure versus IMV, models incorporating SOFA score and either age or PaO2/FiO2 (depending on imputed data usage) reduced heterogeneity to zero. Notably, only the model with imputed data showed statistically significant covariates. For HFNC therapy versus IMV, all models utilized the SOFA score as a covariate, with some including age and APACHE II score. These models achieved zero heterogeneity, with the SOFA score being statistically significant in the model without imputed data or abstract congresses. Similarly, for HFNC failure versus IMV, the APACHE II score and, in some cases, age were used as covariates. All models achieved zero heterogeneity, with the APACHE II score being statistically significant in the model using imputed data. In the meta-regression analysis of the correlation between CPAP and IMV, age and PaO2/FiO2 were used, and zero heterogeneity was achieved across all the models. Age was statistically significant in the two models, while PaO2/FiO2 was significant in all CPAP-failure models. The detailed results of these analyses can be found in Tables 2 and 3.
Table 2NIV modalities against IMV for mortality Meta-regressions modelsNIV modalities and modeln studiesI^2^R^2^CovariantEstimateNIMV with imputed data and abstract congress1792.980Intercept − 0.52Influenza0.11Age0.006PaO2/FiO2 − 0.001NIMV without imputed data and no abstract congress885.560Intercept11.76Influenza − 3.21Age0.006PaO2/FiO2 − 0.17NIMV with imputed data and no abstract congress1593.1917.47Intercept − 0.52Influenza0.11Age0.006PaO2/FiO2 − 0.001HFNC with imputed data and abstract congress1169.6114.3intercept − 2.097Age0.01SOFA0.19HFNC without imputed data and no abstract congress59.0169.45intercept − 15.04APACHEII1.18SOFA0.77HFNC with imputed data and no abstract congress936.2350.02intercept − 2.94Age0.02SOFA0.19CPAP with imputed data and abstract congress80100Intercept0.41Age − 0.01PaO2/FiO2 − 0.01CPAP without imputed data and abstract congress642.430Intercept0.81Age − 0.02CPAP with imputed data and no abstract congress557.900Intercept0.96Age − 0.02Table 3NIV failure against IMV for mortalityModeln studiesI^2^R^2^CovariantEstimateNIMV failure with imputed data and no abstract congress712.1197.01Intercept1.15PaO2/FiO2 − 0.004SOFA − 0.06NIMV failure without imputed data and no abstract congress50100Intercept − 0.05SOFA − 0.15Age0.01HFNC failure with imputed data and no abstract congress70100Intercept8.02APACHEII − 0.61Age − 0.02HFNC failure without imputed data and no abstract congress500Intercept8.06APACHEII − 0.72CPAP failure with imputed data and abstract congress70100Intercept3.81PaO2/FiO2 − 0.03Age − 0.04CPAP without imputed data and abstract congress436.450Intercept4.14PaO2/FiO2 − 0.01Age − 0.4CPAP with imputed data and no abstract congress50100Intercept3.71PaO2/FiO2 − 0.01Age − 0.04
In the case for LOS days in ICU, no model could reduce heterogeneity below 50% in NIMV. The best model presents an I^2^ of 50.70% with imputed data from APACHE II, virus, and age. Without imputed data and 5 studies, only PaO2/FiO2 can reduce I^2^ of 50.90%. Different results were obtained for HFNC with imputed data, in which the inclusion of PaO2/FiO2, APACHE II, and Age can reduce heterogeneity to 0%. Without imputed data, PaO2/FiO2 and APACHE II continue to reduce heterogeneity near to 0%. Models are detailed in Table 4.
Table 4NIV against IMV for LOS days in ICUModeln studiesI^2^R^2^CovariantEstimateNIMV with imputed data650.7770Intercept − 13.50APACHEII0.12Virus2.81Age0.18NIMV without imputed data550.9879.41Intercept0.42PaO2/FiO2 − 0.004HFNC with imputed data40100Intercept − 12.81SOFA0.22Age0.19PaO2/FiO2 − 0.01HFNC without imputed data33.6499.52Intercept − 5.83SOFA − 0.20PaO2/FiO20.03
Although only NRS were used, all outcomes and interventions start with a high (⨁⨁⨁⨁) certainty of evidence given due to the ROBINS-I tool being used. For the risk of bias, the concerns about comparability bias were prevalent due to insufficient adjustments for crucial factors such as age or PaO2/FiO2 in many studies, leading to lowering the GRADE to medium (⨁⨁⨁◯) in NIMV and HFNC therapy followed by intubation and low (⨁⨁◯◯) for the rest of the interventions in mortality. Finally, due to serious imprecision for NIMV, GRADE was lowered to low (⨁⨁◯◯). No problems were detected in inconsistency, indirectness, or imprecision for HFNC therapy followed by intubation. NIMV failure presents serious concerns in indirectness and imprecision, sow certainty was very low (⨁◯◯◯). In the case of LOS days in ICU, NIMV was lowered to medium due to no meta-regression model which could reduce heterogeneity, while HFNC having sensitivity analysis with a different result also lowered the certainty to medium. For a detailed overview, refer to Table 5.
Table 5GRADE evaluation for noninvasive ventilation compared with invasive mechanical ventilation for ARF due to viral pneumonia (assessed with ROBINS-I)InterventionCertainty assessment№ of patientsCertaintyImportance№ of studiesStudy designRisk of biasInconsistencyIndirectnessImprecisionOther considerationsNIVIMVMortality (any cause)NIMV16Observational studiesSerious^a^Not seriousNot seriousSerious^b^All plausible residual confounding would reduce the demonstrated effect10,121/26,992 (37.5%)7396/19,782 (37.4%)⨁⨁◯◯LowCRITICNIMV-F7Observational studiesVery serious^c^Very serious^d^Not seriousVery serious^e^All plausible residual confounding would reduce the demonstrated effect3628/7741 (46.9%)5940/16,017 (37.1%)⨁◯◯◯Very lowCRITICHFNC10Observational studiesVery serious^f^Not seriousNot seriousNot seriousAll plausible residual confounding would reduce the demonstrated effect9761/29,952 (32.6%)5845/16,138 (36.2%)⨁⨁◯◯LowCRITICHFNC-F7Observational studiesSerious^g^Not seriousNot seriousNot seriousAll plausible residual confounding would reduce the demonstrated effect1554/3760 (41.3%)5379/14,581 (36.9%)⨁⨁⨁◯MediumCRITICALCPAP8Observational studiesVery serious^h^Not seriousNot seriousNot seriousAll plausible residual confounding would reduce the demonstrated effect532/1178 (45.2%)265/501 (52.9%)⨁⨁◯◯LowCRITICCPAP-F7Observational studiesVery serious^i^Very serious^j^Not seriousNot seriousAll plausible residual confounding would reduce the demonstrated effect313/486 (64.4%)239/457 (52.2%)⨁◯◯◯Very lowCRITICLength of stay in ICU daysNIMV6Observational studiesNot seriousNot seriousNot seriousSerious^k^All plausible residual confounding would reduce the demonstrated effect476772⨁⨁⨁◯MediumIMPORTANTHFNC4Observational studiesNot seriousSerious^l^Not seriousNot seriousAll plausible residual confounding would reduce the demonstrated effect351528⨁⨁⨁◯MediumIMPORTANTExplanationsa. 8 of the 16 studies used did not adjust or assess important differences in the choice of cohortb. No meta regression model could reduce heterogeneityc. Only 2 of the 7 studies adjusted for or assessed important differences in the choice of cohort. e. 3 of the 7 studies have different effect directionsd. The different magnitude of the effect even in the studies that have the same effect directione. 3 of the 7 studies used did not adjust or assess important differences in the choice of cohortf. Only one study did not adjust or assess important differences in the choice of cohortg. 4 of the 7 studies used did not adjust or assess important differences in the choice of cohorth. 2 of 8 studies used did not adjust or assess important differences in the choice of cohorti. 2 of 7 studies used did not adjust or assess important differences in the choice of cohortj. The different magnitude of the effect even in the studies that have the same effect directionk. 2 of the 6 studies used did not adjust or assess important differences in the choice of cohortl. The sensitivity analysis yields a different result
In this systematic review of patients with virus-induced ARF, most of the studies comparing HFNC, CPAP, and NIV to IMV showed improved survival or no difference in ICU mortality and LOS days in ICU. These results can be contrasted by the exploratory meta-analysis, as their association coefficients revealed significant reductions in mortality or LOS days in ICU. For all NIV-failure patients (NIMV, HFNC, or CPAP), most of the studies did not show an increase in mortality, and similar results were shown in their exploratory meta-analysis. Exploratory meta-regression also explored and explained heterogeneity in all ventilatory strategies, indicating that the severity of the patients affected the results. For LOS days in ICU, only NIMV ventilation reduce those days while HFNC cannot. These results provide valuable insights into the efficacy of different ventilation strategies in managing ARF due to viral infections.
Four meta-analyses and systematic reviews [5, 61–63] have explored the correlation between NIMV as a ventilatory strategy and its impact on mortality. However, our study differs from these previous works in several key aspects. First, the clinical settings assessed varied among these studies. While three of them focused on patients with acute hypoxemic respiratory failure resulting from diverse etiologies, each employing different definitions [61–63], one study specifically examined patients with acute hypoxemic respiratory failure caused by COVID-19, SARS, and MERS [5]. In contrast, our study focused solely on patients with ARF caused by COVID-19, SARS, MERS, or H1N1. Second, the types of studies included in the analysis also differed. Two meta-analyses incorporated only randomized controlled trials (RCTs) [61, 63], while the others used various study designs. In the present study, congress abstracts were included without bias due to insufficient information on their methodology. Consequently, meta-regression models were constructed with and without these abstracts, revealing variations in I^2^ and R^2^ across all NIV modalities. Third, the methods of analysis varied across the studies. Two of them employed pairwise meta-analysis [61, 62], one utilized a network meta-analysis [63], and one, like our study, solely conducted a systematic review [5]. The latter study included several types of interventions [5], while the other two studies assessed only helmet ventilation as an intervention [62]. Additionally, the types of interventions evaluated were not uniform in these studies.
The reduction in mortality observed in NIMV can be attributed to several factors. First, IMV often requires additional interventions such as sedation and vasopressors, which may contribute to higher mortality rates, particularly in critically ill patients with prolonged ICU stays. Additionally, IMV is associated with an increased risk of nosocomial infections, such as ventilator-associated pneumonia [64], making noninvasive ventilation strategies more advantageous for mitigating these complications [65]. However, caution is required when interpreting the findings of studies comparing NIMV failure followed by intubation to initial IMV, as some studies indicate no reduction in mortality for patients requiring intubation after NIMV failure. Further research is needed to identify specific patient factors contributing to increased mortality risk and to refine recommendations for NIMV management.
In the case of HFNC therapy, the positive outcomes in the included studies could be attributed to its ability to generate a low level of positive pressure in the upper airway, possibly improving oxygenation [66]. While some patients who experience HFNC failure have increased mortality, the ROX index has been found to be suitable for predicting HFNC failure in COVID-19 patients with acute hypoxemic respiratory failure [67]. However, many studies in the review did not incorporate the ROX index, and future studies comparing HFNC therapy against IMV should consider its inclusion to improve patient selection accuracy.
Compared with IMV, CPAP was associated with a reduction in mortality in patients with viral-induced ARF. CPAP increases airway pressure, ameliorates arterial oxygenation, increases end-expiratory lung volume, and improves cardiac function [4]. Patients with a PaO2/FiO2 ratio ranging from 100 to 200 can benefit from CPAP, as studies including patients within this range reported positive outcomes. However, due to the limited number of available studies, the safety profile of CPAP in patients with ARF of viral origin requires further investigation.
One of the reasons for the reduction on LOS days in ICU by NIMV and not in HFNC may be attributable to the positive end expiratory pressure (PEEP). The PEEP can improve ventilation-perfusion (VQ) mismatches [68], and those advantages can be seen in less mortality and less LOS days in ICU. Similarly, not having an invasive device may result in fewer infections which translate into fewer LOS days in ICU and less mortality. However, this last hypothesis was not evaluated in the present review with meta-analysis since there were not enough studies reporting nosocomial infections such as ventilator-associated pneumonia.
There are several strengths focused on the search strategy. An example of this was the inclusion of search engines such as Trials.gov and regulatory agencies of Latin American countries for gray literature, which allowed us to reduce the risk of publication bias. However, asymmetries were identified both by funnel plot and by Egger’s test, but there was a bias toward publishing studies with negative results; nevertheless, favorable results were found for all NIV modalities. This tells us that the problem is not in the lack of studies but in their quality with incomplete information.
Our study is subject to several limitations. First, this study relied solely on observational studies due to logistical and ethical challenges in conducting clinical trials during the early stages of the COVID-19 pandemic, potentially introducing biases and overestimating effects. This constraint is reflected in the review’s low to medium GRADE recommendations. For example, a meta-analysis by Masaaki Sakuray comparing NIMV to IMV in ARDS patients yielded mixed results depending on the device used. Additionally, extracting ICU stay data as medians with interquartile ranges posed challenges due to heterogeneity among patients, making assumptions for conversion into means with standard deviations inappropriate. The wide variance in patient numbers across studies also significantly impacts heterogeneity, as observed in sensitivity analyses for each NIV modality. Furthermore, the lack of distinction between ICU and hospital mortality in some studies required an assumption that all deaths occurred in the ICU if not specified.
Moreover, the use of random effects models was necessary due to anticipated high heterogeneity among critical patients, but this limits the generalizability of the meta-analysis results. Subgroup analysis could mitigate this limitation, but defining cutoff points for continuous covariates may introduce biases. Additionally, the absence of mean PaFiO2 ratios in most studies hinders the ability to draw definitive conclusions on the efficacy of NIV or HFNC therapy in ARF patients.
Furthermore, the inclusion of abstracts and missing data limited the possibility of inferential analysis. The imputed covariate data, such as age, PaO2/FiO2, SOFA score, and APACHE II score, were explored in meta-regression models against studies that reported those covariates, along with abstract congress results. However, due to the limitations mentioned, no clinical recommendations can be made based on these meta-regressions. Although imputed data may reduce variance in observational studies, the multitude of NIV modalities could counteract this effect. To address this, data imputation was stratified based on NIV modality. Recommendations for critical care research include the need to report covariates used in meta-regression models. Abstract congresses, due to being on trial, can present different results once they are published. To prevent this bias, multiple meta-analyses and meta-regressions with and without abstract congress were performed, which were used for sensitivity analysis. However, future updates to the present systemic review will be necessary once these abstract conferences are reviewed and published for a more complete and precise evaluation.
The findings of this systematic review support the use of alternative noninvasive oxygenation and ventilation strategies as viable alternatives to conventional respiratory ventilation for managing viral-induced ARF. Although it is essential to interpret these findings with caution given the overall low to medium certainty of the evidence and the quality of being an exploratory meta-analysis, the integration of these modalities as part of the management strategies of these patients could help reduce the utilization of ICU beds, invasive ventilators, and costs in both developed and developing countries.
Additional file Table S1 can be found at Study Summary COVID-19 ventilation.Additional file Table S2 can be found at TableS2.