Machine learning models compared with current clinical indices to predict the outcome of high flow nasal cannula therapy in acute hypoxemic respiratory failure.

Summary

Support Vector Machine models trained on the first 2 hours of HFNC data outperformed ROX-based indices across internal and external datasets. In external validation (n=567), a noninvasive SVM achieved AUC 0.79 (accuracy 73%) vs ROX AUC 0.74; adding ABG features increased AUC to 0.82 with accuracy 83% on MIMIC/eICU.

Key Findings

• Noninvasive SVM model externally validated on 567 AHRF patients achieved AUC 0.79, accuracy 73%, sensitivity 73%, specificity 73%.
• ROX index benchmark showed lower performance (AUC 0.74, accuracy 64%, sensitivity 79%, specificity 60%).
• Including arterial blood gas variables improved SVM external performance to AUC 0.82 and accuracy 83% on MIMIC-IV/eICU.
• Models used only the first 2 hours of HFNC data, enabling early risk stratification.

Clinical Implications

Integrating SVM-based predictions into ward/ICU workflows could prioritize closer monitoring, timely transition to NIV/intubation, and resource allocation; prospective implementation studies and calibration to local data are needed.

Limitations

• Observational design without prospective clinical impact assessment.
• Model interpretability and need for site-specific recalibration; potential dataset shift.

Future Directions

Prospective, pragmatic trials embedding the model in clinical workflows; assessment of impact on time to intubation, ICU LOS, and mortality; fairness and robustness analyses.