Machine learning-based forecast of Helmet-CPAP therapy failure in Acute Respiratory Distress Syndrome patients.

Summary

Using 622 EMR records with 38 features, SVM and neural network models predicted Helmet-CPAP failure in ARDS with high accuracy (up to 95.19%) and strong F1 scores (up to 88.61%). Key predictors included PaO2/FiO2, CRP, and oxygen saturation; performance remained robust after feature reduction.

Key Findings

• SVM achieved 95.19% accuracy and 88.61% F1-score; neural networks achieved 94.65% accuracy and 87.18% F1-score.
• Key features driving predictions were PaO2/FiO2 ratio, C-reactive protein, and oxygen saturation; heart rate, WBC, and D-dimer were secondary.
• Model performance remained high with reduced feature sets (e.g., SVM with 23 features; XGBoost with 13 features).

Clinical Implications

Could support real-time triage for ARDS patients on Helmet-CPAP by flagging likely failure, prompting earlier escalation (e.g., intubation) and optimized resource allocation.

Limitations

• Single-center dataset without external validation, risking overfitting and limited generalizability.
• Retrospective design; no prospective clinical impact analysis or decision-curve evaluation.

Future Directions

External validation across diverse centers, prospective impact studies with clinician-in-the-loop deployment, and calibration/decision-curve analyses to quantify net benefit.