Inclusion criteria were age above 18, mechanical ventilation for more than 48 h, successful 30-min SBT using either T-piece or pressure support ventilation [21], and a LUS ≥ 14 measured at the end of the SBT. Eligibility to SBT was defined according to international conference consensus criteria [22] and undertaken after therapeutic optimisation (see additional file pages 13). Before the spontaneous breathing trial, drainage of large pleural effusion, control of fluid overload, fiberoptic suctioning of atelectasis and bacteriological assessment of distal pulmonary sample were systematically performed.

COPD, tracheostomy, tetraplegia, paraplegia above the 8th thoracic segment, unplanned extubation (accidental or self-extubation), do-not-reintubate order at time of inclusion, refusal to participate, inclusion in another interventional study, major protected, pregnancy, and lack of acoustic window. All patients or their relatives provided written informed consent.

Transthoracic lung ultrasound was performed at the end of the 30-min successful SBT by a senior physician with an expertise in lung ultrasound. Videoslide presentations that were used for providing basic skills in lung ultrasound can be freely downloaded by clicking on the 10 URLs displayed in additional file 3, page 90. As previously recommended [16, 25], 12 regions of interest were examined (see additional file 2, pages 8 and 75 and 76). Each was classified in one of the following categories: 0) normal aeration: horizontal A lines with ≤ 2 vertical B lines; 1) moderate aeration loss: multiple spaced B1 lines and/or localized coalescent B2 lines; 2) severe aeration loss: generalized coalescent B2 lines (pulmonary edema); 3) complete aeration loss: lung consolidation with static and/or dynamic air bronchograms. A value (0, 1, 2 or 3) was attributed to each region examined and the LUS was equal to the sum of the 12 regions examined [16]. Score quotation allowed to differentiate focal from ubiquitous B2 lines (moderate versus severe aeration loss) [26,27,28] (see additional file 3, p 90: performance and LU Score, http://links.lww.com/ALN/C155). High individual risk for PRF was defined as a LUS ≥ 14 at the end of the SBT. Selecting a LUS threshold of 14 is issued from our pilot study published in 2012 [16]. In 86 critically ill patients ready to be extubated, the lung ultrasound score best predicting postextubation respiratory failure (measured at the end of the successful spontaneous breathing trial) was determined. The receiving operating characteristic curve showed a cutoff value for lung ultrasound score > 14 estimated by maximizing the Youden index with an area under the curve of 0.8667.

In patients who successfully passed a SBT, randomisation was performed before extubation following the LUS assessment. Patients were randomly assigned (1:1 ratio) to the LUS intervention group or to the control group. Randomisation was performed centrally through a dedicated, password-protected web-based system, using computer-generated permuted blocks. Within each block, the number of participants allocated to each of the two treatment arms was equal. In the LUS intervention group, the LUS was communicated to the attending physician. Patients with LUS ≥ 14 received prophylactic 2 h HFNO alternating with 1 h NIV during two days whereas patients with LUS < 14 received conventional oxygen. In the control group, the LUS was not communicated to the attending physician and patients received conventional oxygen.

HFNO was delivered using large bore binasal silicon prongs through a heated humidifier (MR850, Fisher & Paykel Healthcare®, Auckland New Zealand). A high flow ranging between 50 and 60 L/min was delivered to meet patient’s comfort. The patient was asked to breathe through the nose with the mouth closed to generate a positive end-expiratory pressure. NIV was delivered through a face mask connected to a ventilator equipped with a MR850 humidifier and providing a pressure support ranging between 5 and 10 cmH2O to obtain a tidal volume of 7–10 ml kg−1 of predicted body weight. The fraction of inspired oxygen (FiO2) and positive end-expiratory pressure level were adjusted to maintain a peripheral capillary oxygen saturation (SpO2) ≥ 92%. Conventional oxygen was delivered by a non-rebreathing face mask using an oxygen flow rate ranging from 10 to 15 L min−1 (mean FiO2 0.37 ± 0.2). In both groups, PRF, defined by predetermined criteria (see additional file 2, pages 15, 67 and 84), was treated by rescue NIV and/or reintubation.

Clinical tolerance, respiratory discomfort and serious adverse events were prospectively assessed and compared between the three modes of respiratory support (see additional file 2, p 86).

The primary outcome was the incidence of PRF within the 48 h following extubation in patients with LUS ≥ 14.

PRF was defined by the occurrence of 2 among the following signs: 1—increase of oxygen flow as O2 mask > 9 L/min to maintain SpO2 > 90% 2—respiratory rate > 30 breaths/min with activation of accessory inspiratory muscles, fatigue, respiratory arrest, major bronchial obstruction 3—hypercapnia > 50 mmHg with pH < 7.35 4—poor hemodynamic tolerance of respiratory origin, heart rate > 120 beats/min, systolic blood pressure > 200 or < 90 mmHg, cardiac arrhythmias, cardiac arrest for respiratory cause 5—poor neurological tolerance of respiratory origin with hypoxic or hypercapnic encephalopathy, agitation, loss of consciousness, confusion, coma, convulsions.

For patients of the LUS intervention group receiving prophylactic NIV + HFNO, PRF was treated by rescue NIV, defined by the increased duration in NIV sessions beyond 8 h per day and the increase in positive end-expiratory pressure to meet oxygen requirements. For patients of the control group, PRF was treated by rescue NIV and/or HFNO. In both groups, death occurring within 48 h following extubation was classified as PRF.

Secondary outcomes were the incidence of PRF and reintubation within the first 7 days following extubation, number of ventilator-free days at day 28 (non-survivors were assigned to have zero ventilator-free day), length of stay in the ICU and mortality in the ICU at day 28 and 90. In both groups, death occurring within 7 days after extubation was classified as PRF.

Using estimates derived from previous data [16], we anticipated that 85% of patients would successfully pass the SBT and that the incidence of PRF in patients with a LUS ≥ 14 would be approximately 65%. We determined that 192 patients would need to be enrolled to provide the trial with 80% power to detect a 30% relative reduction of PRF, at a two-tailed alpha level of 0.05. To account for an incidence of 40% of patients with a LUS ≥ 14 in a similar patient population [16], and a possible incidence of laryngospasm, emergency surgery and cross-over of 10% of patients, enrolment of 650 extubated patients was planned.

The coordinating centre, the statistician and all the investigators remained unaware of the trial-group assignments until the end of the study where data were locked. The trial protocol and statistical analysis plan were evaluated and approved by independant reviewers of the French Ministry of Health when the WIN-IN-WEAN study was submitted for funding to the “Programme Hospitalier de Recherche Clinique Regional 2014” (University Hospital of Poitiers, France PHRCI-14-003).

All statistical analyses were performed in the modified intention-to-treat population, which was prespecified as all randomised patients without those who had to be reintubated after randomisation for laryngospasm or intercurrent surgery. In the modified intention-to-treat population, patients with major protocol violations (patients randomised after mechanical ventilation < 48 h, control patients treated by NIV and/or HFNO, control patients treated by NIV and/or HFNO and LUS intervention patients treated by conventional oxygen) were included in the statistical analysis. For the primary analysis, unavailable data were attributed the worst value of the effect of treatment (maximum bias). In patients with early withdrawal from the study, the number of ICU ventilator-free days at the end of the study were considered. For the analysis of the length of ICU stay and mortality, patients lost for follow up were censored.

The primary outcome was compared between the two groups using an unadjusted chi-square test. Other binary outcomes were tested using an unadjusted chi-square test or Fisher's exact test as appropriate. Results are reported as relative risks with 95% confidence intervals. Adjusted analyses were performed with the use of robust random-effect Poisson generalised linear mixed model regression with robust variance for binary outcomes [29] and linear mixed regression for continuous outcomes, with study site as a random effect. Prespecified covariates were fixed according to clinical relevance. Multicollinearity between variables was assessed by computing the variance inflation factor and using the Farrar–Glauber test. The Akaike information criterion and Bayesian information criterion were calculated and used as model diagnostics to determine how well the model fit improved following addition of covariates. Time-to-event was compared between the two groups using the Kaplan–Meier method. A marginal Cox proportional-hazards model was used to estimate the hazard ratio and 95% confidence interval. The proportional hazard hypothesis was evaluated using the Schoenfeld test and plotting residuals.

No correction for multiple testing was applied in the analyses of secondary outcomes. Complete case analysis was performed for all outcomes. We did not compensate for dropouts. The effects of LUS ≥ 14 and < 14 on outcomes were compared in patients treated by conventional oxygenation using an unadjusted Fisher's exact test. A two-sided p value of less than 0.05 was considered to indicate statistical significance. All analyses were generated with the use of Stata software, version 15.0 (StataCorp).