CS-MORT-6: A mortality score that refines risk within EHR-derived SCAI stages in cardiogenic shock

Background Cardiogenic shock carries high in-hospital mortality, and the ordinal SCAI classification does not rank risk within a stage. Existing scores generally require imaging, procedural, or support-dependent data. Methods We identified 3103 adults with documented cardiogenic shock in MIMIC-IV. Six predictors (lactate, urine output, cardiac arrest, age, blood urea nitrogen, red cell distribution width) were selected by bootstrap LASSO screening and clinical review. CS-MORT-6 was derived as two continuous ridge logistic models (lactate or anion gap) fitted at a 24-h landmark and a 0-to-15-point integer score. The primary outcome was subsequent in-hospital death among patients alive in the ICU at the landmark (n = 2694; 892 deaths). The models were externally validated in 1047 patients from 117 eICU hospitals. Results The integer score's AUROC was 0.727 (95% CI, 0.706-0.747) on cross-validation and 0.759 (0.729-0.790) externally; corresponding values were 0.734 and 0.759 (lactate model) and 0.726 and 0.748 (anion-gap model). Externally, the integer score over-predicted on average (calibration-in-the-large -0.22, slope 1.15). Within EHR-derived SCAI stages, mortality differed by 20 to 47 percentage points between the lowest and highest score tertiles in MIMIC-IV, with a similar pattern in eICU. Adding the anion-gap model to stage increased apparent AUROC by 0.139 (0.116-0.162) in MIMIC-IV and 0.141 (0.101-0.178) in eICU. Compared with BOS,MA2 (654 eICU patients scorable by both), the anion-gap model's AUROC difference was 0.004 (-0.039 to 0.046). Conclusions At 24 h, CS-MORT-6 distinguished mortality risk within EHR-derived SCAI stages and retained moderate discrimination externally. Repeated scoring and real-time performance require prospective evaluation.