The albumin-to-glucose ratio as an independent predictor of 30-day mortality in immunosuppressed patients with pneumonia receiving glucocorticoids: a retrospective multicenter study

Immunocompromised hosts receiving glucocorticoids face high pneumonia mortality, yet there is a lack of rapid, integrative biomarkers for effective risk stratification in this setting. We investigated whether the albumin-to-glucose ratio (AGR) predicts 30-day mortality in this population. This retrospective multicenter cohort included 473 glucocorticoid-treated patients with pneumonia. Data were obtained from the Dryad Digital Repository and used for secondary analysis. Multivariable Cox proportional hazards regression, receiver operating characteristic (ROC) curve analysis, mediation analysis, K-means clustering, and machine learning were used to evaluate AGR’s prognostic performance. Higher AGR independently reduced mortality risk (adjusted HR 0.82, 95% CI 0.73–0.91, P < 0.001), with optimal discrimination at ≥ 5.0 (area under the curve [AUC] 0.713). Platelet count (PLT) mediated only 9.24% of this association. Clustering based on AGR and PLT revealed three distinct phenotypes with graded mortality risks. In the crude model, Cluster 2 (low AGR, low PLT) exhibited a 4.04-fold increase in mortality compared with Cluster 1 (moderate AGR, high PLT) (HR 4.04, 95% CI 2.32–7.03, P < 0.001). However, after full adjustment, the difference between Cluster 2 and Cluster 1 was attenuated and no longer statistically significant (HR 2.02, 95% CI 0.92–4.42, P = 0.079). Boruta feature selection analysis identified AGR as a significant predictor. Among the six machine learning models, the Extratrees model had the best performance, with an AUC of 0.879 for 30-day mortality, comparable to the other models. Of note, these clustering and machine learning findings were derived without internal cross-validation, hyperparameter optimization, or external validation, and should therefore be considered exploratory and hypothesis-generating. AGR showed an independent association with 30-day mortality in immunocompromised patients with pneumonia receiving glucocorticoids. Its discriminative performance was comparable to several conventional severity scores and biomarkers in our dataset. External validation is required before clinical application.