Incremental Prognostic Value of Albumin-Anchored Inflammatory Ratios Beyond a SOFA-Based Model in a Medical ICU: A Retrospective Cohort Study

Background/Objectives: We compared three albumin-anchored ratios calculated from measurements obtained during the first 24 h, C-reactive protein (CRP)-to-albumin (CAR), procalcitonin-to-albumin (PAR) and lactate-to-albumin (LAR), as predictors of intensive care unit (ICU) mortality beyond an established severity model. Methods: Eligibility required an ICU stay reaching 24 h and a complete early work-up; 372 of 1352 screened admissions qualified, with each ratio using the worst component values in that window. Incremental value over a base model (Sequential Organ Failure Assessment (SOFA) score, sex, age-adjusted Charlson index) was assessed by change in the area under the curve (AUC; DeLong test), reclassification metrics (NRI, IDI) and calibration, with alternative baselines and outcomes, multiplicity adjustment, and selection weights from the excluded admissions. Results: Of 372 patients, 248 (66.7%) died. As standalone predictors, CAR (0.671), PAR (0.689) and LAR (0.642) were indistinguishable and inferior to SOFA (0.730). Only CAR significantly improved AUC discrimination (ΔAUC +0.030, p = 0.024; NRI +0.39; IDI +0.042); PAR and LAR did not significantly improve AUC discrimination. Entered together, CAR remained independent (aOR 1.52, 1.15–2.03), while PAR did not (1.20, 0.89–1.63). The cohort was sicker than the 626 patients excluded for incomplete work-up; after weighting, CAR stayed independent (aOR 1.47) but ΔAUC fell to +0.018. The increment was significant against SOFA alone (+0.033) but not against baselines including admission diagnosis or organ support; it remained significant after Benjamini–Hochberg but not Bonferroni adjustment. With in-hospital mortality it was larger (+0.044), but PAR then also retained independence. Conclusions: The first 24 h CAR provided a well calibrated improvement beyond a parsimonious SOFA-based model, but the increment was small and not robust to richer baselines or to multiplicity correction, and no subgroup-specific claim is supported. These hypothesis-generating findings need prospective validation.