Machine Learning-Enhanced Hemoglobin-to-Creatinine Ratio for Prognosis in Critically Ill AMI Patients: A Two-Database Retrospective Cohort Study

Abstract Objective: To investigate whether the hemoglobin-to-creatinine (Hb/Cr) ratio provides independent prognostic value in critically ill patients with acute myocardial infarction (AMI), and further, whether it offers incremental value beyond the Sequential Organ Failure Assessment (SOFA) score. Design: Retrospective observational cohort study with external validation. Setting: ICUs from the MIMIC-IV database (2008–2019, discovery) and the eICU-CRD database (2014–2015, multi-center validation). Patients: A total of 6,182 ICU-AMI patients from MIMIC-IV and 3,761 from eICU-CRD (MICE-analyzable cohorts). Interventions: None. Measurements and Main Results: The primary outcome was in-hospital mortality. A three-tier hierarchical strategy was applied: Model 1 (age+sex) replicated prior literature; Model 2 further adjusted for lactate, mechanical ventilation, and cardiogenic shock; Model 3 added SOFA (and renal indicators in MIMIC-IV) to test incremental value. In MIMIC-IV, Hb/Cr was significantly independently associated with mortality in Model 1 (OR = 1.11, p = 0.002) and remained significant in Model 2 (OR = 1.06, p = 0.031), confirming its basic prognostic value. However, after SOFA inclusion in Model 3, the association was completely attenuated to null (OR = 1.003, p = 0.756). This exact pattern was replicated in eICU-CRD (Model 2 OR = 0.945, p < 0.001; Model 3 OR = 0.989, p = 0.148). SOFA-stratified ΔAUCs were <0.005 across all strata (all p > 0.05), and SHAP confirmed SOFA consistently outranked Hb/Cr in feature importance. Conclusions: Hb/Cr does carry independent prognostic information in ICU-AMI patients under traditional/adjusted frameworks. However, this information is entirely encompassed by the SOFA score. In settings where 24-hour SOFA is already available, Hb/Cr offers no additional clinical utility