Predictive Models for Mortality in Cardiogenic Shock: A Systematic Review and Meta-Analysis

OBJECTIVES: Cardiogenic shock is a time-critical emergency requiring aggressive therapeutic interventions, yet prognostication remains challenging with no consensus on the comparative applicabilities of risk scores. Our study aims to evaluate the capacities of existing risk scores for efficient, effective, and generalizable prognostication of mortality across various patient demographic cohorts. DATA SOURCES: We searched MEDLINE, Embase, and Scopus databases up to June 15, 2024. STUDY SELECTION: We included articles developing, redeveloping, or validating a multivariable model predicting all-cause mortality in adults with cardiogenic shock. DATA EXTRACTION: We pooled area under the curve (AUC) statistics and observed: expected (O:E) ratios as an assessment of discrimination and calibration, respectively. We conducted random-effects inverse-variance weighted meta-analyses on prespecified prediction models (Intra-Aortic Balloon Counterpulsation in Acute Myocardial Infarction Complicated by Cardiogenic Shock IABP-SHOCK II, CardShock, Simplified Acute Physiology Score II, Sequential Organ Failure Assessment, Acute Physiology and Chronic Health Evaluation II, and Survival After Venoarterial Extracorporeal Membrane Oxygenation) to derive pooled estimates of AUCs and aggregate O:Es for each model. Subgroup analyses were conducted to identify sources of heterogeneity. DATA SYNTHESIS: We included 102 studies comprising 126 study cohorts in our final analysis (89,546 patients), with 40 unique prediction models identified. There were no significant differences between the performances of all included scores. However, in terms of absolute pooled values, CardShock score had the highest pooled discrimination (AUC, 0.73 95% CI, 0.70-0.76) and best calibration (O:E, 1.06 95% CI, 0.79-1.41) among other widely used scores. Subgroup analyses were highly variable between studies and again did not reveal a superior prediction model. Unique prediction models developed by authors were found to be superior to existing prediction models, yet lacked generalizability. CONCLUSIONS: Meta-analysis of six included scores revealed no clear "gold standard" prediction model, although the CardShock score had the highest discrimination and most accurate calibration. The degree of variability between studies precluded in-depth assessment of subgroup analysis. Further research into novel risk scores can be conducted to better inform clinicians on their utility.