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

Yu, H; Saffaran, S; Tonelli, R; Laffey, JG; Esquinas, AM; de Lima, LM; Kawano-Dourado, L; Maia, IS; Cavalcanti, AB; Clini, E; Bates, DG

doi:10.1186/s13054-025-05336-4

Respiratory failure & ventilation

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

Crit CarePublished 7 Mar 2025Prognostic model development and external validation study

Read the original article ↗PubMed

Cite

Authors' conclusion

Decision support tools based on SVM models could provide clinicians with more accurate early predictions of HFNC outcome than currently available clinical indices.

Study context

Population: Patients with acute hypoxemic respiratory failure treated with high-flow nasal cannula (HFNC); development at the respiratory ICU of the University Hospital of Modena (2018–2023), external validation in RENOVATE trial (Brazil) and MIMIC-IV/eICU (US)development
Participants: 184 patients

Comparison: Support Vector Machine (SVM) machine learning (ML) model vs clinical indices for predicting HFNC outcome
Outcome: Prediction of HFNC outcome (failure) from measurements made within the first 2 h of treatment
Follow-up: First 2 h of treatment

Methodological detailsmachine learning, internal and external validation

Machine learning
Internal and external validation

Performance

As reported in the abstract

Prediction of HFNC failure, external validation (567-patient dataset)

Test or model	AUC	Sensitivity	Specificity	Accuracy
SVM	0.79	73%	73%	73%
ROX index	0.74	79%	60%	64%

AUC

SVM0.79

ROX index0.74

Sensitivity

SVM73%

ROX index79%

Specificity

SVM73%

ROX index60%

Accuracy

SVM73%

ROX index64%

Further findings

Outcome	SVM with ABG	Modified ROX index
Prediction of HFNC failure with ABG measurements, external validation (MIMIC-IV/eICU dataset)
Accuracy	83%	70%
Sensitivity	84%	63%
Specificity	82%	74%
AUC	0.82	0.65

Reference

Authors: Yu H, Saffaran S, Tonelli R, Laffey JG, Esquinas AM, de Lima LM, et al.
All 11 authors
Yu H, Saffaran S, Tonelli R, Laffey JG, Esquinas AM, de Lima LM, Kawano-Dourado L, Maia IS, Cavalcanti AB, Clini E, Bates DG
Journal: Critical Care
Published: 7 Mar 2025
DOI: 10.1186/s13054-025-05336-4
PubMed: 40055757

Topic: Respiratory failure & ventilation

Comparison

SVM ML model vs clinical indices (ROX and variants, HR/saturation of pulse oxygen (SpO2), SpO2/FiO2, PaO2/FiO2, SOFA, HACOR) for predicting HFNC outcome

Diagnostic definitions

SVM: Support Vector Machine ML model trained on non-invasive measurements
ROX index: respiratory rate oxygenation index
SVM with ABG: SVM model trained also on arterial blood gasses
Modified ROX index: requires PaO2

Key notes

Internal and external predictive performance of SVM superior to all clinical indices across all scenarios tested

Original publication

Machine learning models compared with current clinical indices to predict the outcome of high flow nasal cannula therapy in acute hypoxemic respiratory failure. Crit Care. Published 7 March 2025. doi:10.1186/s13054-025-05336-4

Read the original article ↗

Composed by AI from the published abstract and checked against it. How Briefs are made →