Early dynamic arterial blood gas trajectories to identify clinical phenotypes and predict outcomes in acute hypoxemic respiratory failure: A retrospective cohort study.

BACKGROUND

Acute hypoxemic respiratory failure (AHRF) is a critical syndrome frequently encountered in intensive care, characterized by substantial pathophysiological heterogeneity and poor clinical outcomes. Traditional static risk assessments based on a single time point struggle to capture the dynamic physiological responses to early interventions, thereby limiting precise risk stratification. This study aimed to identify distinct clinical phenotypes in patients with AHRF using dynamic arterial blood gas (ABG) trajectories during early intensive care unit (ICU) stay and evaluate their association with 28-day mortality.

METHODS

This was a single-center retrospective cohort study including patients with AHRF admitted to the ICU of Sichuan Provincial People's Hospital between January 2020 and June 2025. Multiple ABG parameters within the first 24 h of ICU admission-including PaO2/FiO2, PaCO2, pH, HCO₃-, and lactate-were extracted to construct longitudinal trajectory features. Unsupervised K-means clustering was performed to identify dynamic clinical phenotypes. Kaplan-Meier survival analysis and multivariable Cox proportional hazards regression models (adjusted for iatrogenic confounding variables such as the 24-h ABG sampling frequency) were utilized to assess the association between phenotypes and 28-day survival outcomes. The incremental predictive value of the phenotypes relative to the APACHE II score was evaluated using the concordance index (C-index).

RESULTS

A total of 2793 patients with AHRF were included and classified into three distinct clinical phenotypes with unique physiological evolution patterns: Phenotype 1 (Respiratory, n = 882), Phenotype 2 (Metabolic, n = 683), and Phenotype 3 (Stable, n = 1228). Phenotype 2 was characterized by persistent hyperlactatemia and profound oxygenation impairment, exhibiting the highest ICU mortality (18%), 28-day mortality (24%), and hospital mortality (18%) among the three groups. Kaplan-Meier analysis indicated that the 28-day overall survival for Phenotype 2 was significantly worse than that of the other two groups (log-rank P < 0.001). Multivariable Cox regression demonstrated that after adjusting for age, gender, APACHE II score, mechanical ventilation, 24-h ABG sampling frequency, and routine laboratory biomarkers (WBC, creatinine, AST, ALT), Phenotype 2 (Metabolic)remained independently associated with an increased 28-day mortality risk compared with Phenotype 1 (Respiratory, HR = 1.38, 95% CI: 1.12-1.70, P = 0.003). The ABG sampling frequency itself showed no independent association with mortality (HR = 1.01, 95% CI: 0.94-1.10, P = 0.773). Phenotype 3 (Stable) showed a trend toward reduced mortality risk (HR = 0.83, 95% CI: 0.67-1.00, P = 0.104). Integrating dynamic phenotypes into the baseline prediction model significantly improved the C-index for predicting 28-day mortality from 0.657 to 0.671 (P < 0.001).

CONCLUSIONS

Unsupervised clustering based on early dynamic ABG trajectories effectively identifies three distinct clinical phenotypes with divergent prognostic features in patients with AHRF. Phenotype 2 (Metabolic) denotes a substantially higher 28-day mortality risk, whereas Phenotype 3 (Stable) confers a relatively favorable outcome. This dynamic subphenotyping framework provides valuable incremental prognostic information beyond traditional static scoring systems.