Development of a predictive model for high-flow nasal cannula treatment failure in emergency department patients with acute heart failure: a retrospective secondary data analysis

Purpose: This study developed and internally validated a model integrating oxygenation, acid-base, and severity variables for high-flow nasal cannula (HFNC) failure in emergency department (ED) patients with acute heart failure (AHF).Methods: This retrospective secondary analysis included adults with AHF who received HFNC therapy in a tertiary ED between January 2021 and August 2025.Five prespecified predictors-follow-up saturation-to-fraction of inspired oxygen (S/F) ratio, follow-up respiratory rate-oxygenation index incorporating heart rate, norepinephrine use, Sequential Organ Failure Assessment score, and arterial pH-were entered into logistic regression and four machine learning algorithms using a 70:30 stratified split and bootstrap internal validation.Results: Among 201 patients, 33 (16.4%) experienced HFNC failure.In the held-out test set, the five-variable logistic regression model showed an area under the receiver operating characteristic curve (AUROC) of .70 (95% confidence interval [CI], .51-.88), with a bootstrap optimism-corrected AUROC of .78(95% CI, .70-.86) in the full cohort.The five models showed AUROCs ranging from .70 to .81, with overlapping confidence intervals.Per 1-standard deviation increase, the follow-up S/F ratio (odds ratio [OR], 0.20; 95% CI, 0.07-0.58)and arterial pH (OR, 0.55; 95% CI, 0.35-0.86)were independent predictors of HFNC failure, whereas norepinephrine use was not (OR, 2.70; 95% CI, 0.37-19.50).Conclusion: Integrating early oxygenation response with acid-base status may assist in the bedside stratification of HFNC failure risk in AHF.Standard logistic regression using these prespecified parameters may offer an interpretable alternative to machine learning models, although it did not outperform the follow-up S/F ratio alone.