To our knowledge, this is the first study that analysed the independent prognostic value of multiple OI in predicting early failure within 24 h of NIRT applied outside ICU to a large series of patients with ARF due to SARS-CoV-2 Pneumonia; furthermore, it empathises the role of non-oxygenation-related parameters in contributing to their outcome.

The main findings of the study are:

1)effectiveness in predicting NIRT failure varies across the OI and increased over time, being highest after 24 h of respiratory support; conventional and standardized PaO2/FiO2, as well as RI showed significant prognostic value at any time of observation, while O2 gradients predicted NIRT outcome only after 24 h of treatment;

2)SpO2/FiO2 ≤ 300, PaO2/FiO2 ≤ 151,7 and s-PaO2/FiO2 ≤ 160,4 resulted the best predictors of NIRT outcome, respectively at baseline, 12 and 24 h of NIRT;

3) among non-oxygenation parameters, elderly age, male gender, Charlson Index > 3, neurological comorbidities, agitation/confusion, need of sedation, inability to tolerate pronation were independently correlated with NIRT failure, with sedation status being the best prognostic factor.

These findings are coherent with the fact that pathophysiology of hypoxemia in COVID-19 pneumonia is complex and multi-factorial [1, 31,32,33,34]. In the early phase, impairment in alveolar diffusing capacity is the main “driver” of “silent hypoxemia” which is likely to respond to increased FiO2. In the later phases, alveolar consolidations develop, resulting in V/Q mismatch and shunt-related hypoxemia which is more likely to respond to NIRT by means of improving alveolar recruitment [14, 31, 35, 36]. Microvascular thrombosis may worsen hypoxemia shifting V/Q towards higher values [32, 33]. As lung damage progresses, persistent poor oxygenation induces hyperventilation with hypocapnia and tachidyspnea [37].

PaO2/FiO2 ratio is the most widely applied OI which correlates with mortality in ARDS [19] and COVID-19 related hypoxemia [38, 39]. However, it has important physiologic drawbacks. Firstly, PaO2/FiO2 is strongly FiO2-dependent; if the administered oxygen flow is inappropriately increased, PaO2/FiO2 ratio drops and severity of ARF may be overweighted. Secondly, PaO2/FiO2 may underweight the severity of ARF in tachypneic-hypocapnic patients. Thirdly, PaO2/FiO2 cannot provide information on the mechanisms underlying hypoxemia [20]. Despite these patho-physiologic limitations, in our study, PaO2/FiO2 ratio kept a significant prognostic value in COVID-19 related Pneumonia in agreement with other experiences. In our study, PaO2/FiO2 ratio equal or lower than 200 (AUC = 0,654), 181 (AUC = 0,697) and 180 (AUC = 0,776) were independently associated with NIRT failure showing a OR of 2,14, 3,67 and 4,94 respectively at t0, t12 and t24. In the study of Prediletto et al. [40] conducted in 349 COVID-19-related hypoxemic patients (mean PaO2/FiO2 189,4; NIRT used in < 40% of patients), PaO2/FiO2 < 180 significantly predicted failure (AUC 0,742) defined as deaths or need of IMV.

SpO2/FiO2 closely mirrors hypoxemia strata defined by PaO2/FiO2 ratio [35, 36]. Due to the leftward shift of oxy-haemoglobin dissociation curve in hypoxemic and hypocapnic patients, SpO2/FiO2 is less dependent on changes of FiO2, keeping fairly stable values even for larger changes of PaO2 and FiO2. Conversely, SpO2/FiO2 measurement is poorly reliable in shock. Furthermore, it shares the same limitations reported for PaO2/FiO2 being not able to provide information on ventilatory status and mechanisms of hypoxemia. In the earlier phases of COVID-19 with “silent hypoxemia” [35, 36], SpO2/FiO2 may perform as well as ROX index (ratio between SpO2/FiO2 and RR) [41]; this may be explained by the fact that hyperventilation-induced hypocapnic compensation of hypoxemia in COVID-19 is mainly obtained by increase of tidal volume rather than by an increase in RR [40]. Furthermore, SpO2/FiO2 and ROX index at baseline significantly correlated to PaO2/FiO2 in a series of 100 COVID-19 patients with moderate-severe hypoxemia [42].

In our study, SpO2/FiO2 ratio equal or less than 300 (AUC = 0,662) was associated with NIRT failure showing a OR of 3,05 of at t0; this finding turned out to be the strongest predictive OI at that time. These reports are similar to what found in other studies [43, 44]. In a series of 133 severely hypoxemic COVID-19 patients treated with HFNC, Kim et al. demonstrated that SpO2/FiO2 ratio was more accurate than ROX index as predictor of failure, providing the greatest predictivity at 1 h of treatment [43]. In a population of 456 hypoxemic COVID-19 patients managed outside ICU, Cattazzo et al. found that PaO2/FiO2, ROX index and SpO2/FiO2 predicted ETI or death with similar accuracy [44].

s-PaO2/FiO2 well estimates V/Q mismatch in hypocapnic patients [45,46,47] because it adjusts the conventional ratio to the PaCO2 value; thus avoid to underestimate the worsening of lung gas-exchange in hyperventilating patients. In our study, s-PaO2/FiO2 ratio equal or lower than 159 (AUC = 0,615), 151,7 (AUC = 0,672) and 160.4 (AUC = 0,769) were associated with NIRT failure showing a OR of 2,23, 2,56 and 6,88 respectively at t0, t12 and t24; this resulted the strongest predictive OI after 24 h of NIRT. Prediletto et al. [40] demonstrated that s-PaO2/FiO2 predicted death better than conventional PaO2/FiO2; s-PaO2/FiO2 values lower than 170 and 125 best prognosticated failure and mortality, respectively [40].

Oxygen gradients and RI gives information on the patho-physiology of hypoxemia and are influenced by capnia and respiratory effort [48,49,50]. A-aO2 values pathologically increases in case of worsening of V/Q matching and shunt due COVID-19 interstitial and vascular abnormalities [51].

In our study, RI cut-offs were able to predict NIRT outcome failure at any time of observation; its prognostic value increased at 24 h of NIRT. RI cut-off greater than 2,37 at 24 h of treatment was associated with NIRT failure showing a OR of 4,5. Conversely, increased oxygen gradients predicted NIRT failure only after 24 h. In non-COVID-19 pneumonia A-a gradient was a useful indicator of severity and outcome [52,53,54]. In COVID-19 mildly hypoxemic patients, AaDO2 predicted occurrence of severe pneumonia, ICU admission and hospital readmission, but not mortality [55,56,57]. In a series of severely hypoxemic 165 patients started on NIV for COVID-19 pneumonia, the capability of A–a gradient > 430,83 to predict mortality was higher than what found for other arterial blood gas values, including PaO2/FiO2 [17]. Conversely, in a recent study, Singh et al. [18] found that aADO2, adjaADO2 and RI were not sensitive nor specific, with a poor accuracy in predicting mortality in severe COVID-19 pneumonia.

The comparison of our findings with the available scanty published data investigating the prognostic value of OI in NIRT-treated COVID 19 patients is biased by differences in terms of severity of ARF, degree of hypocapnia, types of NIRT, and setting of treatment.

For what non-OI variables concerns, our findings are in agreement with literature data [11, 43, 58, 59]. Accordingly, in our study, age, comorbidities, agitation and delirium were independently associated with NIRT failure, while success in performing pronation had a favourable prognostic value [60]. The rate of NIRT failure observed in our study (< 25%) is consistent with literature data on COVID-19 patients managed with NIRT outside ICU [61, 62]. It should be considered that a substantial proportion of our patients (19.3%) were not candidate to escalation to IMV; in DNI patients NIRT failure was by far greater than in the rest of population (85,9 vs 8,4%) in agreement with the literature [61, 62].

The study has several limitations. Firstly, the retrospective, single center and uncontrolled design of the study may reduce the strength of the results due to potential missed data, such as the smoke habits, and BMI; respiratory comorbities were not entered in the statistical analysis because of very low incidence (COPD, asthma and ILD in < 5% of the study population) in agreement with literature data. However, in the context of the global pandemic, challenges were reported in conducting RCTs. Secondly, the application of NIRT according to a specific internal algorithm may limit the generalizability of our findings compared to centres using other protocols. Thirdly, the analysis did not include RR, has been shown to be a strong predictor of NIRT outcome, especially if applied as a component of the ROX index; the lack of this parameter should have been mitigated by the peculiarity of “silent” COVID-19 ARF where tachypnea and dyspnoea arise usually late in the course of the pneumonia. Fourthly, the incidence of pulmonary embolism and vascular abnormalities as well as CT features and extension of lung infiltrated were not available; however, the aim of the study was to investigate the role of OI in hypoxemia independently on the underlying mechanisms. Finally, we have not been able to include in the prognostic analysis some severity scores (ie. Apache, Saps, Sofa) which are strong predictors of outcome in critically ill patients.

The strengths of this study include the large population analysed and the ability to assess for the first time a wide range of OI at multiple time of NIRT use, to identify the prognostic value of each one in the earlier phases of COVID-19 pneumonia managed outside ICU. This research on OI has potential important implication in non- COVID-19 related hypoxemic patients managed by NIRT in PMICU to select those who are more likely to require ETI and ICU admission.