BACKGROUND: predictors of outcome are useful in COVID-19 to define the best treatment and setting of care. ROX (Respiratory rate-OXygenation) index, has been proposed to predict early failure of non-invasive support strategies in ARF. We evaluated the predictive role of ROX index in identifying treatment failure (i.e., ET or death) in COVID-19 using data from the overall hospitalization. METHODS: retrospective multicentric study on COVID-19 patients from 8 Italian hospitals during the second pandemic wave. RESULTS: we enrolled 227 patients, 157 males (69.2%), median PaO2/FiO2 ratio at admission was 248 [IQR 170-295] and about one third of the population (29.5% N=67) required ET or died during hospitalization. Those who failed were older (70 vs 61 years, p<0.001), and had a higher burden of cardiovascular diseases (74.6% vs 46.3%, p<0.001) than discharged patients. Also, they had a lower PaO2/FiO2 ratio at presentation (median 229 vs 254, p=0.014). There were no statistically significant differences regarding gender, BMI or other comorbidities. In patients who failed treatment, ROX index was significantly higher at presentation and showed a sharp worsening at day 3 and 4, while in patients who succeeded, ROX index was stable over time and, indeed, reduced after 5-6 days. The predictive ability of ROX index improves especially on the third day and we identified its best cut-off value at 8.53 (sensitivity 75%, specificity 68%). Kaplan-Meier curves of the first 3 days from admission confirmed that a ROX index ?8.53 was associated with a higher risk treatment failure (P<0.001). CONCLUSION: ROX index on 3rd day is the best predictor of treatment failure in severe COVID-19 patients.