Predicting High-Flow Nasal Oxygen Failure at 30 Minutes Using a Physiology-Informed Dual-Domain Model

Overview

This prospective, single-center, two-stage translational study develops and validates a physiology-informed dual-domain model for ultra-early (30-minute) prediction of high-flow nasal oxygen (HFNO) failure in patients with acute hypoxemic respiratory failure. The study includes a physiological validation cohort (n=24) to anchor the EIT-derived Flow Index (EFI) as a marker of inspiratory effort, followed by a temporally separated clinical derivation cohort (n=57) and independent validation cohort (n=58). Candidate predictors are screened from 1,328 clinical features. The final dual-domain model integrates persistent physiological burden (baseline PaCO₂ and 30-minute EFI) with short-term dynamic adaptation (ΔRR and ΔSpO₂ over 30 minutes). The model's discrimination is tested prospectively without refitting.

Detailed Description This was a single-center, prospective, two-stage translational study conducted at Ruijin Hospital, Shanghai Jiao Tong University School of Medicine. The study comprised three integrated components: 1. Physiological validation cohort (n = 24) Mechanically ventilated patients with acute respiratory distress syndrome (ARDS) receiving pressure support ventilation underwent simultaneous electrical impedance tomography (EIT) and esophageal pressure monitoring. Measurements were performed at three sequentially adjusted pressure support levels: baseline clinical setting (PSbase), maximal tolerated level (PSmax), and minimal level (PSmin). The EIT-derived Flow Index (EFI) was calculated from the global impedance-time signal. Relationships between EFI and esophageal pressure swing (ΔPes) as well as pressure-time product per minute (PTP/min) were assessed using regression analysis. Changes in EFI across pressure support levels were evaluated by repeated within-subject comparisons. 2. Clinical derivation cohort (n = 57) High-risk adult patients with acute hypoxemic respiratory failure (AHRF) initiated on high-flow nasal oxygen (HFNO) were prospectively enrolled between May 2025 and September 2025. Inclusion required at least one of the following high-risk criteria: PaO₂/FiO₂ ≤ 200 mmHg or FiO₂ ≥ 0.40 to maintain SpO₂ ≥ 92%; respiratory rate ≥ 25 breaths/min; APACHE II score ≥ 12; or bilateral infiltrates on chest imaging. EIT and bedside variables (heart rate, respiratory rate, arterial blood gases, ROX index) were recorded at baseline (HFNO initiation) and at 30 minutes. HFNO failure was defined a priori as clinically meaningful escalation to noninvasive ventilation (NIV) or endotracheal intubation due to sustained hypoxemia, progressive respiratory acidosis, respiratory muscle fatigue, or hemodynamic instability. Within-tier adjustments (increasing flow or FiO₂ without changing support modality) were not considered failure. Patient-level analyses were performed to identify two prespecified domains of early HFNO failure: Persistent abnormality (physiological burden that remained abnormal after accounting for baseline): evaluated by analysis of covariance (ANCOVA) for 30-minute variables adjusted for baseline values. Divergent short-term response trajectory (different evolution between success and failure groups): evaluated by generalized estimating equations (GEE) with time-by-group interactions across baseline and 30 minutes. A multivariable logistic regression model was constructed in the derivation cohort incorporating baseline PaCO₂, 30-minute EFI, ΔRR (change in respiratory rate), and ΔSpO₂ (change in peripheral oxygen saturation). An exploratory reference model using ΔPaO₂ instead of ΔSpO₂ was also evaluated. 3. Prospective validation cohort (n = 58) An independent, temporally separate cohort of patients meeting the same inclusion/exclusion criteria was enrolled between October 2025 and March 2026 (after completion of the derivation cohort). The identical 30-minute reassessment protocol was applied. The prespecified logistic regression equation from the derivation cohort was applied directly without coefficient refitting. Discriminatory performance of the dual-domain model (baseline PaCO₂ + 30-min EFI + ΔRR + ΔSpO₂) was evaluated using area under the receiver operating characteristic curve (AUROC), sensitivity, and specificity. Total enrollment: 164 participants (24 physiological validation + 115 clinical HFNO participants \[derivation 57 + validation 58\] + 25 screened but excluded as detailed in the study flow diagram). The study was approved by the Ruijin Hospital Ethics Committee (Reference Nos. \[2025\]30 and \[2025\]232). All participants provided written informed consent.