Prediction of high-flow nasal cannula failure in COVID-19 acute hypoxemic respiratory failure using combined clinical and metabolomic data: a machine learning study

Background High-flow nasal cannula (HFNC) is widely used as first-line non-invasive respiratory support in acute hypoxemic respiratory failure (AHRF), particularly during the COVID-19 pandemic. However, HFNC failure remains frequent, and delayed intubation is associated with worse outcomes. Accurate early prediction of HFNC failure is therefore essential to guide clinical decision-making. Methods We conducted a prospective observational cohort study including 111 patients with COVID-19-related AHRF treated with HFNC in a single tertiary ICU. Patients were divided into a Training Cohort (March 2020–April 2021; n = 89) and a temporally independent Validation Cohort (May 2021–October 2021; n = 22). A total of 164 variables (25 clinical and 139 metabolomic features) were collected at baseline. Predictive models were developed using Random Forest machine learning with recursive feature elimination and bootstrap resampling. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), with internal validation and temporally independent validation. Results The six-variable model integrating clinical and metabolomic data achieved an AUC of 0.87 (95% CI: 0.79–0.94) in the Training Cohort and 0.91 (95% CI: 0.78–1.00) in the Validation Cohort. A simplified three-variable model (SpO₂/FiO₂ ratio, anthranilic acid, and octanoic acid) achieved an AUC of 0.71 (95% CI: 0.61–0.82) in the Training Cohort and 0.91 (95% CI: 0.77–1.00) in the Validation Cohort. Both models showed improved performance compared with the ROX index in the Training Cohort. Performance in the Validation Cohort should be interpreted cautiously due to limited sample size. Conclusions A predictive model combining clinical and metabolomic data showed promising performance for identifying HFNC failure in COVID-19-related AHRF. A simplified model using only three variables retained comparable performance in the validation cohort, supporting further evaluation of its potential clinical utility. Further validation in larger and more diverse populations is required.