Cardiogenic shock is a time-critical, life-threatening condition characterised by low cardiac output, tissue hypoperfusion and progressive organ failure. Despite advances in treatment and mechanical circulatory support (MCS), mortality remains high. Accurate prognostication is essential to guide clinical decisions, allocate resources, and facilitate communication with patients and families. However, there is currently no consensus regarding which risk score performs best across varied patient populations.
A systematic review was conducted to assess discrimination and calibration of available multivariable mortality prediction models. Discrimination refers to how well a model distinguishes between survivors and non-survivors, while calibration assesses how closely predicted mortality matches observed mortality.
A total of 102 studies comprising 126 cohorts and 89,546 patients were included. Forty unique prediction models were identified. Six commonly used models were selected for quantitative meta-analysis: three general ICU scores (SAPS II, SOFA, APACHE II), two CS-specific scores (CardShock and IABP-SHOCK II), and one MCS-specific score (SAVE).
In terms of discrimination, the CardShock score demonstrated the highest pooled AUC at 0.73 closely followed by SOFA (0.72), and SAPS II and APACHE II (both 0.71). The IABP-SHOCK II score showed an AUC of 0.70, and the SAVE score performed worst at 0.67. Although CardShock ranked highest numerically, statistical comparisons revealed no significant superiority over most other models, except for SAVE, which performed significantly worse.
Regarding calibration, none of the scores consistently over- or under-predicted mortality. The CardShock score again showed the best overall fit with an O:E ratio of 1.06, indicating close agreement between predicted and observed mortality. IABP-SHOCK II tended to underpredict mortality, whereas SAPS II, SOFA, APACHE II and SAVE tended to overpredict mortality, with SAVE showing the greatest overestimation. However, differences were modest, and confidence intervals overlapped, suggesting broadly comparable calibration across models.
Subgroup analyses explored whether model performance varied by age, geographic region, shock aetiology, type of MCS, or study design. Considerable heterogeneity was observed, and no single model consistently outperformed others across all subgroups. CardShock generally showed slightly better discrimination across several contexts, particularly in European cohorts and among patients receiving MCS, but variability limited firm conclusions.
A post hoc analysis evaluated the Society for Cardiovascular Angiography and Interventions (SCAI) shock classification. Although widely used to stage severity, SCAI demonstrated only moderate discriminatory ability (AUC 0.67) and lacks a formal mortality prediction formula, limiting its value as a prognostic tool compared with established scores.
The review also examined newer and locally developed prediction models. Many study-specific or modified scores incorporating additional clinical variables or biomarkers showed higher AUCs within their derivation cohorts. However, these models lacked external validation and generalisability, preventing pooled analysis. Some required laboratory measures or biomarkers that may not be rapidly available in emergency settings, reducing practicality.
Unlike other critical care conditions, such as sepsis or cancer, where disease-specific models clearly outperform general ICU scores, CS-specific models did not demonstrate marked advantages over generic ICU tools. The authors propose that many prognostic factors, such as age, lactate, and renal function, are common to critical illness broadly, which may explain the similar performance across models.
In conclusion, the meta-analysis found no clear “gold standard” mortality prediction model for cardiogenic shock. Although the CardShock score demonstrated slightly better discrimination and calibration, its advantage was small and inconsistent. Established ICU and CS-specific scores perform similarly overall. The authors advocate further development and external validation of newer models, alongside tools that are quick and easy to apply in urgent clinical settings, to better support decision-making in this high-risk population.
Source: Critical Care Medicine
Image Credit: iStock
