This multicentre prospective observational cohort study enrolled consecutive patients admitted to ten Italian COVID-19 centres distributed throughout the country between February 1th, 2021, and July 31st, 2021. We included in the study all adult patients with a diagnosis of SARS-CoV-2 infection, confirmed by standard procedure, reverse transcriptase-polymerase chain reaction (RT-PCR) by a nasal or oropharyngeal swab). Patients <18 years old, those in need of intubation at inclusion, or those with other known diagnoses of interstitial lung diseases, primitive or metastatic lung cancer, heart failure or pregnancy were excluded.

All anamnestic, demographic and other clinical data were recorded and all patients performed arterial blood gas analysis upon the admission. Additional detail is provided in an online data supplement.

The patients were followed up with a daily check-up of all medical records until discharge or death. Noninvasive respiratory support needed during hospitalization was categorized into four groups: first, no support; second, nasal cannulas, venturi mask (VM) or non-rebreathing mask (NRM); third, high flow nasal cannula (HFNC); finally, continuous positive airway pressure (CPAP) and/or pressure support non-invasive ventilation (NIV). Oxygen and ventilatory support to acute respiratory failure was provided according current recommendations.16 PaO2/FIO2 ratio (P/F) was calculated based of the ratio between arterial O2 pressure and oxygen inspiratory fraction administered. In patients requiring low-flow oxygen system supplementation FiO2 was estimated according Shapiro formula. [Fraction of Delivered O2= 0.20 + (0.04 x L/min O2)].17

The clinicatrials.gov register number is NCT04871685 (Home - ClinicalTrials.gov).

All patients underwent LUS at admission to the hospital. LUS was performed by expert operators in each centre who were trained in dedicated courses according to the recommendation of Italian Society Ultrasonography in Medicine and Biology (SIUMB).18 They were blinded to the clinical data or underwent LUS before HR-CT if that was performed. LUS was performed adopting the 12-region model, 6 on each side, with each hemithorax divided into anterior, lateral, and posterior areas (delimited by the anatomical landmarks represented by axillary lines) and each area into upper and lower segments. Other technical factors have been described in an on-line data supplement.

LUS artefacts have been categorized as follows: A-line, horizontal artefacts observed in normal lungs; B-lines: vertical artefacts in a variety of patterns including focal, confluent or “light beam”, which is a lucent, band-shaped and vertical b-line, that move rapidly with sliding; consolidations, single or multifocal, with occasional mobile air bronchograms and white lung. Finally, the state of pleural line was assessed. The findings were classified according to the following scoring method with scores ranging from 0 to 3 (Fig. 1):

Figure 1.

Lung ultrasound imaging related to LUS score (0–3). Score 0: normal A-lines with a continuous and regular pleural line. Score 1: multiple separated B-lines. Score 2: coalescent B-lines pattern with alterations of the pleural line.Score 3: consolidation area and possibly a large white lung artefact.

(0.34MB).

Score 0: normal A-lines with a continuous and regular pleural line.

Score 1: multiple separated B-lines.

Score 2: coalescent B-lines pattern with alterations of the pleural line.

Score 3: consolidation area and possibly a large white lung artefact.

The total score was computed as the sum, which could range from 0 to 36.19

Before the beginning of the study, all sonographers looked at the clip models with different patterns to reduce possible bias in the interpretation of the images. In the case of doubtful scores, the clips were collectively discussed and assigned a shared score.

The primary endpoint of the study was the association of LUS score with in-hospital mortality. The secondary endpoint was the assessment of LUS score association with highest degree of respiratory support the patient required during the hospitalization.

Categorical data are expressed as numbers and percentages, while continuous variables are expressed as either medians and interquartile ranges or means and standard deviations, based on their distribution, which was assessed graphically and by the Shapiro–Wilk test. The presence of missing data is reported. The endpoint was in-hospital mortality, assessed either from data at discharge or the death certificate. The median follow-up time was calculated by the inverse Kaplan–Meier procedure. Univariable and multivariable logistic regression models were performed to evaluate the associations between in-hospital mortality and exposure variables. Odds ratios and 95 % confidence intervals (OR and 95 % CI) were calculated for all models. Nonlinear associations of LUS scores were tested using multivariable fractional polynomial (MFP) models and restricted cubic spline (RCS). The multivariable model was constructed using the best subset selection algorithm (selection with evaluation of all 2k possible models, where k is the number of variables under analysis), with subsequent choice of the best model according to the Bayesian information criterion (BIC). In addition to this selection, the model was enriched with additional variables of relevant clinical importance and known in the literature as prognostic factors. Internal validation was carried out through bootstrap resampling, with 800 repetitions, to evaluate the optimal LUS score alone (adjusted by age and sex) and the proposed model using the bsvalidation package in STATA.20

A p value <0.05 was considered statistically significant. All analyses were performed using statistical software STATA v16 (StataCorp. 2019. College Station, TX: StataCorp LLC).

Additional detail on the statistical analysis is provided in the online data supplement.

The minimum sample size was calculated using the pmsampsize statistical package from STATA.21 The evaluation was performed assuming a binary outcome of mortality of approximately 25 % and a c-statistic of the model of 0.85 considering a maximum of 15 covariates. The calculation resulted in a minimum sample size of 398 patients.

The study was approved by the local ethics committee (University of Campania Luigi Vanvitelli Prot. 0,009,416/I, Naples, Italy, 21–03–15) and was in accordance with the 1976 Declaration of Helsinki and its later amendments. All patients gave their written consent.

A total of 1007 patients are enrolled in the study and evaluated in the statistical analysis. The patients are mainly males (62.4 %), with a mean age of 66.3 years (SD 14.8) and a median duration of hospitalization of 17 days [inter-quartile range (IQR): 10–28 days]. All clinical and laboratory characteristics at admission are reported in Table 1. P/F median for each Center is described in Appendix Table 1.

The LUS score shows a median value of 24 with an IQR of 13 to 30. Appendix Fig. 1 shows the distribution of the scores with modal peaks at multiple scores of 12. Additionally, Appendix Fig. 2 reports the distribution of LUS scores across each centre. The values of the other respiratory variables (PaO2, PaCO2, lactates, SpO2 and P/F ratio) are shown in Table 1. The LUS scores show a strong negative correlation (rho: −0.71) with the P/F ratio and moderate negative correlations with PaO2 and SpO2 (rho: −0.34 and −0.32, respectively) (Appendix Fig. 3).

The variable of respiratory support is used to assess the clinical course of the patients. Although the transition between different support needs is a temporal variable, the brevity of the transitions suggested that we evaluate the worst of the respiratory possibilities as the outcome, thus using an ordinal logistic model. LUS scores show a significant association with respiratory support [Odd ratio (OR): 1.19, 95 % confidence interval (CI): 1.17–1.21; p < 0.001], with an increasing risk of needing more invasive respiratory support as the LUS scores measured on patient arrival increased. Oxygen support is needed in 811 (80.5 %) patients and are subsequently categorized into no support (19.1 %), NC+MV+NRM (32.5 %), HFNC (14.3 %) and CPAP+NIV (34.1 %).

Regarding the LUS score values, the patients who require more respiratory support have higher LUS score values on arrival at the hospital (LUS score median and IQR: no support = 10, 4–16; NC, MV, NRM= 22, 12–24; HFNC= 25, 20–30; CPAP, NIV= 30, 24–36). In addition, a similar trend is evident when assessing LUS score values in patients with different P/F ratios (median LUS score and IQR: P/F < 200 = 26, 24–32; P/F 200–300 = 16, 12–24; P/F > 300= 10, 4–13). (Appendix Table 2). The number of patients requiring oro-tracheal intubation, invasive ventilation and ICU admission is 50 patients (5 %). Baseline respiratory support in patients undergoing IOT for NIV-failure is reported in Appendix Table 3. The median baseline LUS score in patients who are subsequently intubated is 32 (IQR: 28–36). (Appendix Table 4).

During the observation period, 265 in-hospital mortality events are recorded, with a cumulative incidence of approximately 26.3 % and a median follow-up time of 21 days (Fig. 2). The univariable logistic regression analysis for in-hospital mortality is reported in the supplementary table (Appendix Table 5). LUS scores show a significant association with in-hospital mortality (OR: 1.18, 95 % CI: 1.16–1.21; p < 0.001). MFP and RCS show no evidence of a nonlinear association. In the multivariable analyses, depending on the best subset selection algorithm and evidence from the literature, we adjusted the logistic regression model for age, sex, P/F ratio, SpO2, lactate, hypertension, chronic renal failure, diabetes, and obesity (Table 2). Additionally, oxygen support, although not measured at baseline but during the observation period, is included in the model for a sensitivity analysis. Even after adjustment, the LUS score continues to be an important prognostic factor (OR: 1.11, 95 % CI: 1.07–1.14; p < 0.001). Data about in hospital mortality according to respiratory support are showed in appendix Table 6. Results of multivariate analysis including main confounding factors are showed in Appendix Table 7.

Figure 2.

The Kaplan-Meier curve for the cumulative incidence of in-hospital mortality of the study population.

(0.07MB).

Internal validation of the LUS score alone (adjusted by age and sex) and the proposed model result in Brier scores of 31.6 and 39 and c-statistic values of 0.84 and 0.88, respectively (the calibration slope and coefficients from the model adjusted by bootstrap shrinkage are reported in Fig. 3 and Table 3 respectively). Considering the complete model and classifying as positive (death prediction) a patient with logistic regression predicted value >=0.5 [Pr(D) >=0.5], we obtained a sensitivity and specificity of 64.2 % and 91.1 %, with a positive predictive value of 73.3 % and a negative predictive value of 87 %.

Figure 3.

The calibration slope displays, in the form of a scatterplot, the comparison between the predicted outcome risk (x-axis) and the observed outcome risk (y-axis).

(0.29MB).