Predictive Value of Respiratory Parameters for Non-invasive Ventilation Failure in Acute Respiratory Failure: A Systematic Review

Non-invasive ventilation (NIV) is first-line therapy in several forms of acute respiratory failure (ARF), but failure is common and delayed intubation carries excess mortality. The best bedside predictors, and when to measure them, are unclear. Following PRISMA 2020, four databases were searched up to 25 June 2026. Peer-reviewed primary studies reporting an association between a respiratory parameter and NIV failure in adults with ARF were eligible. Risk of bias was assessed using the Quality in Prognosis Studies (QUIPS) tool, and a narrative synthesis was performed. Study selection is presented in a PRISMA flow diagram. Twenty studies from 10 countries (~11,700 patient-episodes) covered chronic obstructive pulmonary disease (COPD) exacerbation, de novo hypoxemic failure, pneumonia and acute respiratory distress syndrome (ARDS), including approximately 8,700 hypoxemic and 3,000 hypercapnic episodes. Failure ranged from 13% to 70%, indicating marked heterogeneity. Values before NIV discriminated poorly; those at one to two hours performed well. A partial pressure of arterial oxygen to fraction of inspired oxygen ratio (PaO2/FiO2) below 146 mmHg at one hour predicted failure (odds ratio 2.51, 95% CI 1.45-4.35), while at or below 200 mmHg, the adjusted odds ratio was 4.26 (95% CI 1.62-11.16). Respiratory rate alone lost predictive value under NIV, whereas high expired tidal volume (>9-9.5 mL/kg) and unrelieved inspiratory effort marked high risk. The composite heart rate, acidosis, consciousness, oxygenation, and respiratory rate (HACOR) score reached areas under the curve of 0.85 (95% CI 0.84-0.87) to 0.94 at one to two hours, falling to 0.71-0.78 externally. Eleven studies were at low risk of bias, seven moderate, and two high. Respiratory parameters predict NIV failure most accurately when reassessed one to two hours after initiation, and composite indices such as HACOR outperform single variables. As the evidence is observational and prone to incorporation bias, confidence is moderate.