Sex-Specific Evaluation and Redevelopment of GRACE Score in NSTE-ACS
- There are major differences in the presentation, risk stratification, and management of NSTE-ACS between men and women.
- This study assesses the sex-specific performance of the GRACE score, which was developed to risk-stratify patients presenting with NSTE-ACS.
- The study confirmed that the GRACE 2.0 underestimated in-hospital mortality in women and tended to incorrectly stratify them to the low- to intermediate-risk group. The machine learning–based GRACE 3.0 score improved the predictive performance in both sexes.
What is the sex-specific performance of the GRACE (Global Registry of Acute Coronary Events) 2.0 score and can it be improved?
The GRACE score was developed to risk-stratify patients presenting with non–ST-segment elevation acute coronary syndrome (NSTE-ACS). This study re-evaluated the performance of the GRACE 2.0 score in a sex-disaggregated manner and redeveloped a new score using machine learning–based approaches to account for individual heterogeneity and nonlinear relationships between sex and conventional risk factors. The authors used data from nationwide cohorts from the United Kingdom (UK) (n = 400,054), which consisted of the training and internal validation cohorts for GRACE 3.0, in addition to an external validation cohort from Switzerland (n = 20,782). The primary outcome was in-hospital mortality, with death at 6 months and 1-year post-admission for NSTE-ACS as secondary outcomes. The discriminatory performance of the GRACE score was assessed for men and women separately using the area under the receiver operating characteristic curve (AUC). A supervised machine learning approach (ensemble learning) based on a tree-based algorithm was used to derive the GRACE 3.0 score using the UK cohort and validated in the Switzerland cohort.
Overall, women presenting with NSTE-ACS had marked differences in GRACE components compared to men in all cohorts. Women were also less likely to receive coronary angiography and undergo early invasive therapy than men. The AUC of the GRACE 2.0 score to predict in-hospital death was 0.86 (95% confidence interval [CI], 0.86-0.86) in men and 0.82 (95% CI, 0.81-0.82) in women. The GRACE 2.0 score underestimated in-hospital mortality risk in women and tended to incorrectly stratify them to the low- to intermediate-risk group, for which the score does not indicate early invasive treatment. This is likely due to the differential weighting of the clinical features of GRACE 2.0 between men and women. Accounting for sex differences, GRACE 3.0 showed superior discrimination and good calibration with an AUC of 0.91 (95% CI, 0.89-0.92) in men and 0.87 (95% CI, 0.84-0.89) in women in the external validation cohort. Sex-specific GRACE 3.0 risk estimates led to reclassification of women towards the high-risk group and of men towards the low- to intermediate-risk group.
The GRACE 2.0 score underestimates in-hospital mortality in women with NSTE-ACS. The machine learning–based GRACE 3.0 score improved predictive performance for both sexes.
The GRACE score has long been used to risk-stratify patients presenting with NSTE-ACS and guide early invasive therapy, despite it being derived from predominantly male patient populations and increasing awareness of its differing discriminative performance between sex. This study uses data of over 400,000, confirms the sex-based differences in discriminatory capacity of the GRACE score, and derives a new version of the score using a machine learning–based approach. This approach accounts for the nonlinear and sex-specific relationships between the GRACE components and in-hospital mortality, improving sex-based risk discrimination for both men and women. Systematic approaches to risk assessment that eliminate gender biases are crucial in order to address the persistent and clinically impactful sex-based differences in the management of patients with NSTE-ACS. Agnostic, machine learning–based methods, which account for unappreciated relationships between variables, represent the future of modeling risk—as long as the input (the data) is of high quality and adequate for the intended purpose. Otherwise, garbage in, garbage out!
Clinical Topics: Acute Coronary Syndromes, Invasive Cardiovascular Angiography and Intervention, Noninvasive Imaging, Prevention, Interventions and ACS, Interventions and Imaging, Angiography, Nuclear Imaging
Keywords: Acute Coronary Syndrome, Coronary Angiography, Hospital Mortality, Machine Learning, Primary Prevention, Myocardial Infarction, Risk Assessment, Risk Factors, Sex Characteristics, Supervised Machine Learning, Women
< Back to Listings