Early lung ultrasound score changes predict the failure of non-invasive respiratory supports in acute hypoxemic patients: a multicenter prospective observational study

Abstract

Background

To determine whether lung ultrasound (LUS) may early predict the failure of non-invasive respiratory support (high-flow nasal cannula-HFNC, continuous positive airway pressure-CPAP, non-invasive ventilation-NIV) in hypoxemic patients.

Methods

In this prospective multicenter international observational study, we enrolled patients undergoing non-invasive treatments for hypoxemia (PaO₂/FiO₂ < 300 mmHg). LUS, PaO₂/FiO₂ and ROX index were assessed before (baseline) and 2 h after treatment start. Regional/global LUS aeration scores were computed (4 degrees of loss-of-aeration: 0-normal to 3-severe loss of aeration) in 6 regions per hemithorax (2 anterior, 2 lateral, 2 posterior). Failure was defined as need of respiratory support’s escalation within 48 h (HFNC to CPAP to NIV, any support to intubation/ECMO).

Results

We studied 100 patients (age 70 [57–76] years; female sex 39%; supports: 13 HFNC, 68 CPAP, 19 NIV); the overall rate of treatment failure was 22%. At the baseline, clinical and ultrasound parameters were similar in failing and non-failing patients; after 2 h, failing patients had lower PaO₂/FiO₂. (149 mmHg [124–201] vs. 200 [171–243]; p = 0.001), lower ROX index (7.8 [4.9–9.2] vs. 10.9 [7.9–13.8]; p = 0.003) and higher lateral (3.0 [1.0–6.0] vs. 1.5 [0.0–3.0]; p = 0.047), antero-lateral (4.0 [1.0–9.0] vs. 2.0 [0.0–4.0]; p = 0.027) and global (13.0 [8.0–17.0] vs. 10.0 [7.0–13.0]; p = 0.036) LUS aeration scores. No improvement in lung aeration was observed in failing patients within the initial 2 h of treatment (global LUS score variations 0.0 [-2.0–1.0] vs. -3.0 [-5.0 – -2.0]; p < 0.001). ROX index and antero-lateral/global LUS scores’ variations were independent predictors of failure. AUCs for treatment failure were: 2-hour ROX index 0.71 [0.58–0.84], 2-hour PaO₂/FiO₂ 0.73 [0.60–0.85], global LUS score variations 0.73 [0.62–0.89]. A combined clinical-ultrasound score (ROX-US) showed AUC of 0.82 [0.73–0.91]. A ROX-US≥1 identified the success of the treatment with sensitivity 95% and specificity 50%; a ROX-US≥2 identified the success of the treatment with sensitivity 45% and specificity 96%.

Conclusions

Changes in LUS aeration scores induced by 2 h of non-invasive respiratory support help early predict the risk of treatment failure. LUS score improved only in responders and was an independent predictor of failure.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13054-025-05791-z.

Background

Hypoxemic acute respiratory failure (ARF) is one of the most frequent causes of admission to intensive care unit (ICU). It requires oxygen administration delivered by different invasive or non-invasive respiratory support techniques [1]. Although their respective role is debated, non-invasive respiratory support – i.e., high-flow nasal cannula (HFNC), continuous positive airway pressure (CPAP) and non-invasive ventilation (NIV) – are worldwide applied first-line treatment of hypoxemic patients [2, 3]. Recent data from a large meta-analysis suggest that non-invasive respiratory supports reduce work of breathing and improve gas exchanges, finally leading to lower risk of intubation and mortality [4]. The importance of non-invasive respiratory support in hypoxemic patients has been further emphasized during coronavirus disease 2019 (COVID-19) outbreak [5]. However, the failure of non-invasive supports is associated to higher mortality, as patients intubated after a non-invasive trial may be exposed to delayed intubation and patient self-induced lung injury [6, 7]. Early detection of failure of non-invasive respiratory support is therefore essential to prevent any delay in invasive mechanical ventilation but remains a challenge. A score based on simple clinical parameters as respiratory rate (RR), peripheral oxygen saturation (SpO₂) and fraction of inspired oxygen (FiO₂) (ROX index) has been proposed to detect the failing patients undergoing HFNC [8], although with moderate accuracy in the early phases [9].

Lung ultrasound (LUS) is increasingly used in critically ill patients [10]. A quantitative approach based on number and type of findings allows the computation of the LUS aeration score (then simply called LUS score), a reliable quantification of lung aeration [11–13]. This quantitative approach has been used to assess and monitor lung aeration in ARF [14, 15]; moreover, being a bedside non-irradiating tool showing quick changes in response to treatment and ventilation setting, it has been applied in patients under invasive mechanical ventilation to assess the effects of PEEP titration [16–19], prone positioning [20, 21], fluid challenges [22] or a 1-hour spontaneous breathing trial [23, 24]. This led to a recent experts’ consensus defining the technical aspects and multiple clinical applications of quantitative LUS in intensive care [25]; it is here recognized that limited knowledge is available on the use of quantitative LUS for the monitoring of non-invasive respiratory supports in adults since a limited number of study in very specific populations (e.g., CPAP in COVID-19 patients during the pandemic outbreak [26, 27] or HFNC in trauma patients [28]) have been done.

This prospective observational cohort study was conducted to evaluate whether LUS score and its variations within 2 h after the institution of non-invasive respiratory support may early discriminate patients prone to treatment failure.

Methods

Aim of the study, design and setting

This was a multicenter, prospective observational study conducted in three university ICUs: Anestesia e Rianimazione 1, Fondazione IRCCS Policlinico San Matteo, Pavia (Italy); Anestesia e Rianimazione, Ospedale Maggiore della Carità di Novara (Italy); Service de Réanimation Chirurgical, Centre Hospitalier Universitaire, Dijon (France). Local ethics committees approved the study and patients’ consent was obtained according to local regulations.

Population

All consecutive patients between June 2018 and November 2021 receiving non-invasive respiratory support for hypoxemia (Partial Arterial Oxygen Pressure (PaO₂)/Fraction of Inspired Oxygen (FiO₂) < 300 mmHg) per clinical indication were enrolled; following a first submission and peer-review, the population was enlarged with new enrolments from August 2023 to February 2025. Non-invasive respiratory supports included HFNC, CPAP and NIV (facemask or helmet). Exclusion criteria were age < 18 years, any limitation to a complete LUS examination (extended thoracic dressing, subcutaneous emphysema, open wounds, pain, burns, etc.), incomplete or low-quality LUS images and previous enrolment in this same study.

Definition of failure of non-invasive respiratory support

Non-invasive respiratory support failure was defined as any escalation in respiratory support within 48 h from non-invasive treatment initiation. Within non-invasive supports, treatment escalation was defined as the transition across different types of respiratory supports as follows (e-Figure 1): HFNC to CPAP to NIV. This definition was due to: (1) higher expected PEEP levels transmitted through CPAP vs. HFNC - particularly through helmet CPAP which was the reference technique in most centers]; (2) the further delivery of positive inspiratory support in NIV vs. CPAP [4, 29–32]. All non-invasive supports were considered failed in case of patient intubation and/or ECMO.

Fig. 1.

Diagnostic performance for the failure of non-invasive respiratory supports 2 h after the support onset of (A) PaO₂/FiO₂ and ROX index; (B) Global and antero-lateral LUS scores; (C) variations of global and antero-lateral LUS scores; (D) ROX-US score. ROX index: [SpO₂/FiO₂]/RR; LUS: lung ultrasound; ROX-US: post-hoc score combining cut off values of ROX index and global or antero-lateral LUS score; AUC: area under the curve

The final decision to escalate respiratory support – based on classical criteria such as SpO₂ < 85% despite high FiO₂; lack of improvement in signs of respiratory muscles fatigue; changes in mental status, rendering the patient unconscious or unable to tolerate non-invasive respiratory support; hypotension, despite adequate volume challenge, the use of vasopressors, or both; copious secretions that could not be adequately cleared; lack of improvement in pH or in partial pressure of carbon dioxide (PaCO₂) [33, 34] – was made by the attending physician, who was blinded to LUS findings and recorded the most relevant reason.

Patients dropped out if intubation was necessary during the follow-up for any other reason than respiratory failure (e.g., indication to surgery under general anesthesia).

Protocol

Ultrasound, arterial blood-gas analyses, and clinical parameters were collected right before the non-invasive respiratory support was introduced baseline) and after 2 h as depicted in e-Figure 1. Variations of quantitative variables (delta) were computed as follows: values after 2-hours – baseline values.

Ultrasound examination

Ultrasound examinations were conducted with CX-50 Philips (Amsterdam, Netherlands), Vivid-iQ GE Healthcare (Chicago, Illinois, USA) and Toshiba Xario 200 (Tokyo, Japan) machines.

A complete examination was performed in six standard regions per hemithorax (superior and inferior regions in the anterior, lateral, and posterior fields as defined by sternum, anterior and posterior axillary lines) [25]. The examination was performed in transversal scan with a 9 MHz linear probe, the focus pointed on the pleural line and the depth set at 6–8 cm, adjusted to be at least twice the depth of the pleural line. A phased array probe was used in case of important subcutaneous thickness or if a lung consolidation was visualized.

In each region, lung aeration was graded by LUS as recommended [25]: score 0 – normal aeration: A lines or < 3 well-spaced B-lines; score 1 – mild loss of aeration: ≥3 well-spaced or coalescent B-lines and/or subpleural consolidations occupying ≤ 50% of the visualized pleura; score 2 – moderate loss of aeration: ≥3 well-spaced B-lines or coalescent B-lines and/or subpleural consolidations occupying clearly > 50% of the visualized pleura; score 3 – severe loss of aeration: predominant tissue-like pattern with thickness ≥ 2.5 cm corresponding to lobar/hemilobar consolidations. Subpleural consolidations are defined as echo-poor regions applied to the pleura of at least 5 mm diameter. The score was computed according to the worst ultrasound pattern observed during the respiratory cycle.

Global, anterior, lateral, antero-lateral and posterior LUS scores were computed as the sum of all, anterior, lateral, antero-lateral and posterior regions, ranging therefore 0–36, 0–12, 0–12, 0–24 and 0–12, respectively.

The sonographers were either recognized expert operators (SM, FM, BB) or trainees having concluded a 25-examination supervised training focused on LUS score assessment [35]. Image clips were stored for secondary offline analysis by expert physicians (SM, FM).

Clinical data

The following data were collected:

• Characteristics of patients: age, sex, height, weight, Body Mass Index (BMI); Simplified Acute Physiology Score II (SAPS2), cardiovascular and respiratory comorbidities, time and cause of ICU hospitalization, reason for indication to non-invasive respiratory support;

• Type and device of respiratory support (HFNC, CPAP, NIV, by mask/helmet); FiO₂, airflow, PEEP and pressure support if adequate; eventual main reason of failure and device used for treatment escalation;

• Arterial blood gas analysis prior to initiation of support and after 2 h (PaO₂, PaCO₂, pH, PaO₂/FiO₂);

• Respiratory monitoring prior to initiation of support and after 2 h: respiratory rate (RR) and ROX index computed as (SpO₂/FiO₂)/RR.

Statistical analysis

A sample size of 100 patients was calculated to detect a difference of 3 ± 6 points in the early global LUS score or its change from baseline between failing and non-failing patients, with a two-sided alpha level of 0.05 and 90% power. A failure rate of 30% was assumed based on available literature [3, 36, 37].

Median and interquartile range (IQR) were used for the quantitative variables, number and percentages for the categorical ones. Normal distribution was assessed by Shapiro-Wilk test.

Primary endpoint was to compare early global LUS score (after 2 h of support) and its variations from baseline values in failing vs. non-failing treatments; a Wilcoxon/Mann-Whitney U-test was used, as for the other quantitative variables. Comparisons between baseline values and 2-hour values among the categorical variables were evaluated with the Pearson chi-square test. Comparison between respiratory supports was performed by Kruskal Wallis test and Dunn test for pairwise comparison. Patients with missing data were excluded.

Analyses were repeated considering intubation the only criterion for failure and stratified in post-extubation distress vs. other causes of hypoxemia.

Secondary endpoints were:

1. To assess and compare diagnostic performance of clinical and ultrasound parameters. Receiving Operator Characteristic (ROC) curves and Area Under the Curve (AUC) were computed with relative 95% Confidence Interval (CI); ROC curves were compared with DeLong test. The optimal cut off point was identified by the Youden index.

2. To evaluate the associations between failure and clinical/ultrasound features: odds ratios (ORs) and corresponding 95% CI were calculated. A multivariate regression analysis was also performed; the independent variables were chosen according to statistical significance (univariate analysis, p ≤ 0.05) and to clinical judgment. Age and sex were used as adjusting variables. Separate models were computed for global and antero-lateral LUS scores to avoid nested data.

A post-hoc analysis was performed combining the best cut-off points obtained by Youden index to combine ultrasound and ROX index into a new clinical-ultrasound score called ROX-US.

P-value ≤ 0.05 was considered significant (two-sided). All the analyses were conducted with STATA/SE, version 14.2.

Results

Population

Overall, 116 patients were enrolled; 8 were excluded for: intubation for reasons other than respiratory failure (N = 1: 1 hematemesis), incomplete stored ultrasound examination (N = 3), low-quality ultrasound examination (N = 4) (e-Figure 2). Enrolment lasted longer than expected because the COVID-19 pandemic imposed an interruption of research activities from February to August 2020 and from October 2020 to June 2021 and the new enrolments from August 2023 to February 2025 overlapped with new ongoing protocols on non-invasive ventilation. 100 patients (age 70.0 [57.0–75.5.0.5] years, male sex 61%, SAPS2 40.0 [30.0–51.0] points, BMI 25.9 [23.4–30.6] kg/m²) were considered for final analysis; additional population’s features are displayed in e-Table 1. Enrolment was distributed as follows among the participating centers: Pavia 71, Novara 16, Dijon 13 patients. All the patients presented hypoxemic ARF (median PaO₂/FiO₂ 138 [106–182] mmHg); the underlying cause was post-extubation distress (i.e., patients who were not directly extubated to that support and developed subsequently a respiratory distress – 25), atelectasis (20), cardiogenic pulmonary edema (17), ARDS (16), community-acquired pneumonia (14), aspiration pneumonia (4), thoracic trauma (2), COPD decompensation (1), viral interstitial pneumonia (1). The primary diagnosis for ICU admission was mainly ARF (43.0%); consistently, 39% of patients were enrolled the day of admission to ICU. Details on the post-extubation distress population are displayed in the Supplementary material 1.

Non-invasive respiratory treatments

At the baseline, the 100 patients were supported by low-flow oxygen therapy (57), HFNC (41) and helmet CPAP (2); details are provided in e-Table 1. The respiratory supports chosen for enrolment were 13 HFNC, 68 CPAP (64 with helmet and 4 with mask), and 19 NIV (7 with helmet and 12 with mask). The treatment was started 1.0 [0.0–4.0] days after the admission to ICU. 22 patients (22.0%) failed: 5 of the 13 HFNC (HFNC failure rate 38.5%): 3 were escalated to helmet CPAP and 2 to invasive mechanical ventilation; 13 of the 68 CPAP (CPAP failure rate 19.1%): all were escalated to invasive mechanical ventilation, 4 of the 19 NIV (NIV failure rate 21.1%): 3 were escalated to invasive mechanical ventilation and 1 to ECMO). None of the patients changed support level before the 2-hour assessment. The main cause of failure was persisting hypoxemia (68.2%); details are displayed in Table 1.

Table 1.

Characteristics of the 100 enrolled patients

100 patients undergoing non-invasive respiratory support
Indication to non-invasive support – No. (%):  Risk of post extubation failure  Atelectasis  Cardiogenic pulmonary oedema  ARDS  Community-acquired pneumonia  Aspiration pneumonia  Thoracic trauma  COPD decompensation  Viral interstitial pneumonia	100 25 (25.0) 20 (20.0) 17 (17.0) 16 (16.0) 14 (14.0) 4 (4.0) 2 (2.0) 1 (1.0) 1 (1.0)
Days from admission to enrolment – median [IQR], days	1.0 [0.0–4.0]
Type of respiratory support after enrolment – No. (%):  HFNC  CPAP:  o Mask  o Helmet  NIV  o Mask  o Helmet	100 13 (13.0) 68 (68.0) 4 64 19 (19.0) 12 7
Failure, No (%):  HFNC  CPAP  NIV	22 5 (22.7) 13 (59.1) 4 (18.2)
Main reason defining the failure, No. (%):  Persisting hypoxemia  Inadequate airway control  Persisting severe dyspnea  Hypercarbia	22 15 (68.2) 3 (13.6) 3 (13.6) 1 (4.6)
Respiratory support introduced after failure, No. (%):  Invasive Mechanical ventilation  CPAP  ECMO	22 18 3 1

ARDS: Acute Respiratory Distress Syndrome; COPD: Chronic obstructive pulmonary disease; HFNC: high-flow nasal cannula; CPAP: continuous positive airway pressure; NIV: non-invasive ventilation; ECMO: extra-corporeal membrane oxygenation

Ultrasound and clinical findings

Ultrasound and clinical findings prior to initiation of support and after 2 h are displayed in Table 2 and e-Figure 3.

Table 2.

Ultrasound and clinical features before and after the introduction of the non-invasive respiratory support

	Overall (n = 100)	Successful (n = 78)	Failed (n = 22)	P value
BASELINE: Before respiratory support
LUS aeration scores:  Anterior  Lateral  Antero-lateral  Posterior  Global	1.0 [0.0–3.0] 3.0 [1.5–5.0] 5.0 [2.0–8.0] 9.0 [7.0–10.0] 14.0 [10.0–17.0]	1.0 [0.0–3.0] 3.0 [1.0–5.0] 5.0 [2.0–8.0] 9.0 [7.0–10.0] 13.5 [10.0–17.0]	1.0 [0.0–2.0] 4.0 [3.0–6.0] 4.5 [3.0–8.0] 8.0 [8.0–10.0] 14.5 [11.0–17.0]	0.594 0.211 0.622 0.936 0.640
PaO2/FiO2, mmHg	138 [106–182]	140 [112–188]	131 [84–159]	0.149
PaCO2, mmHg	36.8 [32.0–41.5]	37.0 [33.0–41.1]	34.0 [29.1–41.5]	0.207
pH	7.4 [7.4–7.5]	7.4 [7.4–7.5]	7.4 [7.4–7.5]	0.412
RR, bpm	22.5 [19.0–30.0]	22.0 [18.0–28.0]	25.0 [21.0–34.0]	0.079
ROX index	8.3 [5.5–11.9]	8.7 [5.9–12.5]	6.9 [3.2–10.6]	0.054
EARLY: After 2 h of respiratory support
LUS aeration score:  Anterior  Lateral  Antero-lateral  Posterior  Global	0.0 [0.0–2.0] 2.0 [0.0–4.0] 2.0 [0.5–6.0] 8.0 [5.0–9.0] 10.0 [7.0–14.5]	0.0 [0.0–1.0] 1.5 [0.0–3.0] 2.0 [0.0–4.0] 8.0 [5.0–9.0] 10.0 [7.0–13.0]	1.0 [0.0–4.0] 3.0 [1.0–6.0] 4.0 [1.0–9.0] 8.0 [7.0–9.0] 13.0 [8.0–17.0]	0.068 0.047 0.027 0.255 0.036
PaO2/FiO2, mmHg	189 [150–240]	200 [171–243]	149 [124–201]	0.001
PaCO2, mmHg	36.0 [32.4–42.1]	36.4 [33.0–42.0]	35.3 [30.0–42.8]	0.456
pH	7.4 [7.4–7.5]	7.4 [7.4–7.5]	7.4 [7.4–7.5]	0.037
RR, bpm	22.0 [18.0–26.0]	21.5 [18.0–25.0]	25.0 [21.0–28.0]	0.017
ROX index	9.6 [7.1–13.3]	10.9 [7.9–13.8]	7.8 [4.9–9.2]	0.003

LUS: lung ultrasound; PaO₂: arterial oxygen partial pressure; PaCO₂: arterial carbon dioxide partial pressure, FiO₂: fraction of inspired oxygen; bpm: breaths per minute; SpO₂: Peripheral blood oxygen saturation; RR: respiratory rate; ROX index: [SpO₂/FiO₂]/RR). In bold: significant p values

Data are displayed as median [IQR]

Prior to initiation of non-invasive respiratory support, failing and non-failing patients showed similar clinical and ultrasound parameters.

After 2 h, the failing patients showed lower PaO₂/FiO₂ (149 [124–201] vs. 200 [171–243] mmHg; p = 0.001), lower SpO₂/FiO₂ (187 [144–200] vs. 226 [181–248]; p = 0.002), higher RR (25.0 [21.0–28.0] vs. 21.5 [18.0–25.0]; p = 0.017) and lower ROX index (7.8 [4.9–9.2] vs. 10.9 [7.9–13.8]; p = 0.003).

The absolute values of lateral, antero-lateral and global LUS scores were higher in failing patients, corresponding to a more severe loss of aeration (lateral: 3.0 [1.0–6.0] vs. 1.5 [0.0–3.0], p = 0.047; antero-lateral 4.0 [1.0–9.0] vs. 2.0 [0.0–4.0], p = 0.027; global 13.0 [8.0–17.0] vs. 10.0 [7.0–13.0], p = 0.036).

Table 3 displays the variation of clinical and ultrasound parameters from baseline to after 2 h. A larger improvement in PaO₂/FiO₂ was observed in non-failing patients (53.6 [24.4–94.6] vs. 27.5 [−1.3–67.7]; p = 0.039). The failing patients showed no variation of LUS aeration scores; this corresponded to no impact of the non-invasive respiratory support on lung aeration. On the contrary, a decrease in all LUS scores was observed in non-failing patients, corresponding to an improvement in lung aeration; the difference in LUS scores variation between failing and non-failing patients was statistically significant for all the scores (p < 0.05).

Table 3.

Variation of ultrasound and clinical features 2 h after the introduction of the non-invasive respiratory support (after 2 h – baseline values)

	Overall (n = 100)	Successful (n = 78)	Failed (n = 22)	P value
LUS aeration scores: Anterior Lateral Antero-lateral Posterior Global	0.0 [−1.0–0.0] −1.0 [−2.0–0.0] −1.0 [−3.0–0.0] −1.0 [−2.0–0.0] −3.0 [−5.0 – −1.0]	0.0 [−2.0–0.0] −1.0 [−2.0–0.0] −2.0 [−3.0–0.0] −1.0 [−2.0–0.0] −3.0 [−5.0 – −2.0]	0.0 [0.0–1.0] 0.0 [−1.0–0.0] 0.0 [−1.0–1.0] 0.0 [−2.0–0.0] 0.0 [−2.0–1.0]	< 0.001 0.045 < 0.001 0.065 < 0.001
PaO2/FiO2, mmHg	51.3 [19.3–86.3]	53.6 [24.4–94.6]	27.5 [−1.3–67.7]	0.039
PaCO2, mmHg	0.3 [−2.0–2.4]	0.2 [−1.9–2.1]	0.4 [−2.7–3.4]	0.513
RR, bpm	−2.0 [−5.0–1.5]	−2.0 [−5.0–2.0]	0.0 [−3.0–1.0]	0.502
ROX index	−0.9 [−0.3–3.1]	1.1 [−0.2–4.2]	0.2 [−0.8–1.8]	0.128

Variations (delta) were computed as after 2 h – baseline values. LUS: lung ultrasound; PaO₂: arterial oxygen partial pressure; PaCO₂: arterial carbon dioxide partial pressure, FiO₂: fraction of inspired oxygen; RR: respiratory rate; ROX index: [SpO₂/FiO₂]/RR). In bold: significant p values. Data are displayed as median [IQR]

The variation of global LUS score with different supports is displayed in e-Figure 4; different supports had different impact on lung aeration: median variations of the global LUS aeration score were significantly higher with CPAP (−3.0 [−5.0 – −1.5]) when compared to HFNC (0.0 [−3.0–0.0]; p = 0.002) or NIV (−2.0 [−4.0 – −1.0]; p = 0.029).

Diagnostic performance

The AUCs values for the identification of failure of non-invasive respiratory support for ultrasound and clinical parameters are displayed in Table 4. At the baseline, neither clinical nor ultrasound parameters showed good diagnostic performance (p = n.s.); ROC curves for baseline findings are displayed in e-Figure 5.

Table 4.

Diagnostic performances of clinical and ultrasound parameters for the prediction of the failure of non-invasive respiratory supports

		AUC	95%CI	P value	Optimal cut off	Sensitivity	Specificity
Baseline	ROX index	0.635	0.496–0.774	0.057	5.0	0.87	0.41
	SpO2/FiO2	0.620	0.488–0.753	0.074	234.2	0.31	0.91
	PaO2/FiO2	0.601	0.453–0.749	0.181	86.5	0.96	0.32
	Antero-lateral LUS score	0.534	0.399–0.670	0.619	2.5	0.82	0.29
	Global LUS score	0.533	0.389–0.676	0.655	21.5	0.18	0.91
After 2 h	ROX index	0.708	0.581–0.836	0.001	10.7	0.51	0.86
	SpO2/FiO2	0.709	0.589–0.828	0.001	195.8	0.65	0.73
	PaO2/FiO2	0.725	0.598–0.851	< 0.001	154.3	0.83	0.68
	Antero-lateral LUS score	0.654	0.514–0.793	0.031	3.5	0.64	0.69
	Global LUS score	0.646	0.505–0.788	0.043	12.5	0.55	0.74
	Antero-lateral LUS score variations	0.734	0.610–0.862	< 0.001	−0.5	0.68	0.73
	Global LUS score variations	0.753	0.618–0.886	< 0.001	−0.5	0.59	0.88
	Antero-lateral ROX-US	0.800	0.717–0.884	< 0.001	0.5	0.90	0.55
	Global ROX-US	0.815	0.726–0.904	< 0.001	0.5	0.95	0.50

PaO₂: arterial oxygen partial pressure; PaCO₂: arterial carbon dioxide partial pressure, FiO₂: fraction of inspired oxygen; RR: respiratory rate; ROX index: [SpO₂/FiO₂]/RR); LUS: lung ultrasound (Variations were computed as after 2 h – baseline values); ROX-US: post-hoc score combining cut off values of ROX-index and Global LUS score. In bold: significant p values

After 2 h of treatment, the diagnostic performance improved for ROX index (AUC 0.708, 95%CI 0.581–0.836, p = 0.001), antero-lateral LUS score variations (AUC 0.734, 95%CI 0.610–0.862, p < 0.001) and global LUS score variations (AUC 0.753, 95%CI 0.618–0.886, p < 0.001); no significant difference was found between clinical and ultrasound parameters’ performance (p > 0.9). ROC curves for early findings are displayed in Fig. 1.

Diagnostic performance did not change when adjusted by baseline support (see Supplementary material 1).

In a multivariate analysis, when adjusted per age and sex, both ROX index and variations of global LUS score (e-Table 2) or antero-lateral LUS score (e-Table 3) resulted to be independent predictors of failure of non-invasive respiratory treatments when measured 2 h after of the onset of the respiratory support.

In a post hoc analysis, a clinical-ultrasound score based on the identified cut-off values (ROX-US: 1 point if ROX index > 10.7 and 1 if any improvement in LUS score) had AUC 0.815 [0.726–0.904] for the success of respiratory support. ROX-US performed significantly better than global LUS score alone (p < 0.01) and ROX index (p = 0.035). In our population, 15 patients had ROX-US = 0 and 11 failed (73.3%); 49 had ROX-US = 1 and 10 failed (20.4%); 36 had ROX-US = 2 and 1 failed (2.7%). A global ROX-US≥1 identified the success of the treatment with sensitivity 94.9% and specificity 50.0% (best cut off point with Youden index); a ROX-US≥2 identified the success of the treatment with sensitivity 44.9% and specificity 95.5%.

Diagnostic performance at predicting need for intubation

Considering intubation the only criterion for failure, 20 patients (20.0%) failed (of the 3 HFNC escalated to helmet CPAP, 1 was intubated 24 h later). After 2 h of support, the failing patients showed lower PaO₂/FiO₂ (149 [118–196] vs. 200 [169–242] mmHg; p = 0.002), higher respiratory rate (25.0 [22.5–29.0] vs.21.0 [18.0–25.0]; p = 0.006), lower ROX index (7.3 [4.7–9.0] vs. 10.9 [8.0–13.8]; p = 0.001) and higher absolute values of lateral, antero-lateral and global LUS scores (lateral: 3.0 [1.0–7.0] vs. 1.5 [0.0–3.0], p = 0.040; antero-lateral 4.0 [1.5–10.0] vs. 2.0 [0.0–4.5], p = 0.002; global 13.0 [8.5–18.5] vs. 10.0 [7.0–13.0], p = 0.045). LUS scores variations in failing vs. non-failing patients were different for anterior (0.0 [0.0–1.0] vs. 0.0 [−2.0−0.0]; p < 0.001), lateral (0.0 [−1.0−0.0] vs. −1.0 [−2.0−0.0]; p = 0.037), antero-lateral (0.0 [−1.0−1.0] vs. −2.0 [−3.0−0.0]; p < 0.001) and global (0.0 [−2.0−1.0] vs. −3.0 [−5.0- −1.5]; p < 0.001) scores.

After 2 h, the diagnostic performance was similar to the baseline for ROX index (AUC 0.737, 95%CI 0.607–0.867, p < 0.001) and PaO₂/FiO₂ (AUC 0.728, 95CI 0.560–0.860; p = 0.001) while it significantly improved for ultrasound parameters (antero-lateral LUS score variations AUC 0.752, 95%CI 0.622–0.880, p < 0.001; global LUS score variations AUC 0.765, 95%CI 0.629–0.902, p < 0.001). ROX-US had AUC 0.823 [0.730–0.919], performing better than global LUS score alone (p = 0.004) but not than ROX index (p = 0.094).

Details and ROC curves are displayed in Supplementary material 1.

Discussion

The main results of the present study are: (1) Baseline LUS scores are not predictive of non-invasive respiratory support failure, (2) Early changes in LUS scores after 2 h of treatment are independent predictors of failure, (3) Failing patients show no improvement in lung aeration and 3) A combined clinical-ultrasound approach may improve the early identification of the acute hypoxemic patient failing non-invasive respiratory support.

A major challenge in managing hypoxemic patients with non-invasive respiratory support is avoiding delayed endotracheal intubation which may expose the patient to prolonged self-induced lung injury and have a negative impact on outcome [6, 7]. However, guidelines are unclear about which are the most adequate monitoring tools and timing for patients’ early assessment [38].

In patients with hypoxemic ARF, different values of ROX index at different timing accurately predicted success of HFNC [8, 9] with AUC similar to our results. A single ROX index cut-off might not work in all ARF etiologies [39], all timing [9, 40], and all devices (face vs. helmet CPAP) [41]. Moreover, the accuracy of ROX index in this last context was low in the early phases, being acceptable only after 24 h [41]. This may explain why in our mixed population, we found a new cut-off value. Repeated ROX index measurement might be useful to increase its accuracy [9]; however, we did not find significant variations after 2 h of respiratory support.

Known predictors of NIV failure are older age [42], SAPS2 [42, 43], PaO₂/FiO₂ [44, 45] and tidal volume [43, 44]. A bedside scale named HACOR (heart rate, acidosis, consciousness level, oxygenation, RR) has been proposed [34] and validated [45] to predict NIV failure in hypoxemic patients, with accuracy 81.8% after one hour [45]. In our population, RR and PaO₂/FiO₂ were significantly different after 2 h in failing patients.

Integrating lung imaging in the assessment of non-invasive respiratory supports seems intuitive; being bedside, repeatable and sensible, ultrasound is the ideal instrument. LUS gained a leading role in ARF in the last years [10]; a quantitative approach based on the computation of the LUS aeration score allows a reliable quantification of regional and global aeration [11, 12] and has been applied for the assessment of aeration modification induced by treatments or procedures in patients under invasive mechanical ventilation [13–24].

Studies on quantitative LUS applied to non-invasive respiratory supports are limited. A first retrospective study was performed in blunt chest trauma patients with HFNC: no improvement was observed after 72 h in patients requiring intubation. However, the earliest assessment (within 12 h) was poorly informative. In COVID-19 patients, a single LUS examination before the onset of non-invasive respiratory support allowed prediction of its failure [26, 27]. However, in the pandemic context the scores were extremely high, the LUS examination was performed in a restraint number of regions [26], and the failure rate was also high, thus limiting the generalizability of these results. A PEEP-titration based on LUS in COVID-19 patients in CPAP was also proposed, showing an impact on outcome and confirming the interest of LUS in the optimization of the non-invasive ventilation setting [46].

In our population, the baseline LUS scores – as PaO₂/FiO₂ and ROX index – were similar in failing and non-failing patients. Clinical pictures of ARF with similar severity may differently respond to non-invasive respiratory support according to the underlying etiology [47]. The modifications induced by 2 h of non-invasive support are more informative: the failing patients are more hypoxemic and present a lower ROX index. Accordingly, no variations in LUS scores are observed, corresponding to no impact of the non-invasive support on lung aeration. In successful patients, a significant improvement is observed mostly in global and antero-lateral LUS scores. Our data suggest that an early ultrasound examination showing no improvement in LUS score should orient the physician to early treatment escalation, avoiding a prolonged and potentially harmful non-invasive ventilation trial. Lung aeration variation is predominant in antero-lateral fields, similarly to what found in previous studies on weaning from mechanical ventilation [23, 24]. While it is recommended to perform an initial complete assessment [25], the second assessment after the start on non-invasive respiratory support could be limited to antero-lateral fields, making it faster and more feasible.

The diagnostic performance of LUS score variations, clinical parameters and ROX index are similar. Being ROX index and LUS scores independent predictors of failure, each looking at the patient from a different perspective, a combination of the two could lead to a higher diagnostic accuracy. Post-hoc scores combining ROX index and LUS score showed higher AUCs than ROX index or LUS score alone; in practice, a patient presenting both clinical and imaging response (ROX-US = 2) is identified as successful with 95.5% specificity. However, this clinical-ultrasound approach was the result of a post-hoc analysis that requires external validation in future studies before clinical adoption.

This study presents limitations. First, the prevalence of failure was lower than expected for power analysis. Second, the timing to define early detection of failure was arbitrary (2 h); some literature describes a variation of physiological parameters already after 1 h [34]. However, no clear indication on the timing for non-invasive respiratory monitoring is available in the most recent ESICM guidelines [38] and experts suggest reassessing patients under non-invasive support every 1–2 h [48]. Third, HFNC and CPAP are frequently considered equal non-invasive supports, leading to a different definition of failure. While a study clearly demonstrating that CPAP or NIV after HFNC improves outcome is still missing, trials did show a potential benefit for CPAP/NIV over HFNC [4, 29–32], mainly in hypoxemic patients and mainly if a helmet is used [49, 50]. A physiological study confirmed the possibility to deliver higher PEEP levels with helmet CPAP, when compared to HFNC, with a positive impact on oxygenation and dyspnea [29]. Accordingly, larger trials on hypoxemic acute respiratory failure showed a positive impact on outcome (mortality and need of intubation) when comparing helmet CPAP vs. HFNC [32] or helmet NIV vs. HFNC [31]. A large meta-analysis also confirmed that helmet respiratory supports outperform HFNC in reducing mortality and intubation rate [4]. While a persisting hypoxemia under HFNC is a failure for this device, the patient may still have the potential to positively respond to other NIV devices that allows higher PEEP level and/or pressure support, as for helmet CPAP and NIV. This corresponds to the clinical practice of many centers; ignoring such an escalation in these contexts would mean leaving out an event that reflects a critical step in the patient’s clinical trajectory and a signal of worsening patient’s status. However, if this escalation is delivered too late, its positive effects are limited [51]. This study aims to early identify the patients failing non-invasive respiratory support with a given device in order to provide prompt escalation; in our cohort, the 3 patients failing HFNC were consistently switched to helmet CPAP.

Fourth, the decision to escalate the support was clinical; different clinical approaches may lead to different timing to treatment escalation in different clinical settings. Finally, power analysis was not computed to test the performance of a new clinical-ultrasound score; this is a suggestion derived from our data that need to be confirmed prospectively in a wider population.

Conclusions

Changes in LUS aeration scores after 2 h of non-invasive respiratory support early predict treatment failure. LUS score improved only in responders and its variation was an independent predictor of failure. Further studies should confirm if a combined clinical-ultrasound approach may improve early identification of non-invasive respiratory supports’ failure.

Abbreviations

ICU

Intensive care unit

HFNC

High flow nasal cannula

CPAP

Continuous positive airway pressure

NIV

Non-invasive ventilation

LUS

Lung ultrasound

PaO₂

Partial Arterial Oxygen Pressure

FiO₂

Fraction of Inspired Oxygen

ECMO

Extra-corporeal membrane oxygenation

BMI

Body mass index

RR

respiratory rate

IQR

Interquartile range

AUC

Area under the curve

Author contributions

All authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by all authors. The first draft of the manuscript was written by Silvia Mongodi, Francesco Mojoli, Rosanna Vaschetto and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

Institutional funding.

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This study was performed in line with the principles of the Declaration of Helsinki. Approval was granted by the Ethics Committee of Fondazione IRCCS Policlinico S. Matteo (May 25th, 2018/No. P-20180046310). Informed consent was obtained from all individual participants included in the study.

Consent for publication

Not applicable.

Competing interests

FM received fees for lectures from GE Healthcare, Hamilton Medical, SEDA SpA, outside the present work. SM received fees for lectures from GE Healthcare, outside the present work. A research agreement is active between University of Pavia and Hamilton Medical. RV received an honorarium for a lecture from Intersurgical. DLG has received payments for travel expenses by Getinge, Draeger and Hamilton, personal fees by Draeger, and research grants by Fisher and Paykel and GE. The other authors declare no conflicts of interest.