Nonlinear association between stress hyperglycemia ratio and mortality in AHRF: validation through consensus clustering and supervised machine learning models.

Purpose: Acute hypoxemic respiratory failure (AHRF), a life-threatening condition in critically ill patients, is associated with poor clinical outcomes. Although the stress hyperglycemia ratio (SHR) may reflect acute metabolic stress, its prognostic role in AHRF remains unclear. This study investigated the association between SHR and mortality risk in patients with AHRF. Methods: We retrospectively analyzed 2514 patients with AHRF from the MIMIC-IV database. Restricted cubic spline analysis was used to assess the potential non-linear association between SHR and 28-day all-cause mortality. Kaplan-Meier survival analysis and Cox proportional hazards models were used to evaluate mortality risk. Secondary predictive analyses assessed the incremental value of SHR beyond SOFA and evaluated 101 candidate machine-learning models or model combinations. Consensus clustering was used as an exploratory approach to identify data-driven clinical phenotypes. Results: In the MIMIC-IV cohort, 28-day all-cause mortality increased most clearly among patients with elevated SHR. In the primary parsimonious Cox model, the highest SHR quartile was associated with increased 28-day all-cause mortality compared with the lowest quartile (HR = 2.972, 95% CI: 2.108-4.189, P < 0.001). Restricted cubic spline analysis suggested a possible non-linear association, with a data-derived inflection point around 0.929 in this cohort. Event-distribution and extreme-value sensitivity analyses indicated that the association between elevated SHR and mortality was not driven solely by extreme SHR values. In secondary predictive analyses, SHR improved discrimination beyond SOFA, and the final CoxBoost + SuperPC model achieved an optimism-corrected C-index of 0.803 in the training cohort and a C-index of 0.826 in the independent validation cohort. Consensus clustering identified three exploratory clinical phenotypes with different severity profiles and mortality risks. Phenotype-stratified analyses suggested a clearer SHR-mortality association in Phenotype II, although formal interaction testing with dichotomized SHR was not statistically significant. Conclusion: SHR was associated with 28-day all-cause mortality in patients with AHRF and may serve as a potential prognostic marker for risk stratification. The most consistent evidence supported an increased mortality risk among patients with elevated SHR, whereas the possible nonlinear pattern, particularly in the lower SHR range, requires further validation. Exploratory phenotype-specific findings should be interpreted cautiously and validated in independent cohorts.