Sex inequalities in cardiovascular risk prediction

Abstract

Aims: Evaluate sex differences in cardiovascular disease (CVD) risk prediction, including use of (i) optimal sex-specific risk predictors and (ii) sex-specific risk thresholds.

Methods and results: Prospective cohort study using UK Biobank, including 121 724 and 182 632 healthy men and women, respectively, aged 38-73 years at baseline. There were 11 899 (men) and 9110 (women) incident CVD cases (hospitalization or mortality) with a median of 12.1 years of follow-up. We used recalibrated pooled cohort equations (PCEs; 7.5% 10-year risk threshold as per US guidelines), QRISK3 (10% 10-year risk threshold as per UK guidelines), and Cox survival models using sparse sex-specific variable sets (via LASSO stability selection) to predict CVD risk separately in men and women. LASSO stability selection included 12 variables in common between men and women, with 3 additional variables selected for men and 1 for women. C-statistics were slightly lower for PCE than QRISK3 and models using stably selected variables, but were similar between men and women: 0.67 (0.66-0.68), 0.70 (0.69-0.71), and 0.71 (0.70-0.72) in men and 0.69 (0.68-0.70), 0.72 (0.71-0.73), and 0.72 (0.71-0.73) in women for PCE, QRISK3, and models using stably selected variables, respectively. At current clinically implemented risk thresholds, test sensitivity was markedly lower in women than men for all models: at 7.5% 10-year risk, sensitivity was 65.1 and 68.2% in men and 24.0 and 33.4% in women for PCE and models using stably selected variables, respectively; at 10% 10-year risk, sensitivity was 53.7 and 52.3% in men and 16.8 and 20.2% in women for QRISK3 and models using stably selected variables, respectively. Specificity was correspondingly higher in women than men. However, the sensitivity in women at 5% 10-year risk threshold increased to 50.1, 58.5, and 55.7% for PCE, QRISK3, and models using stably selected variables, respectively.

Conclusion: Use of sparse sex-specific variables improved CVD risk prediction compared with PCE but not QRISK3. At current risk thresholds, PCE and QRISK3 work less well for women than men, but sensitivity was improved in women using a 5% 10-year risk threshold. Use of sex-specific risk thresholds should be considered in any re-evaluation of CVD risk calculators.

Keywords: Biomarkers; CVD risk prediction; Pooled cohort equations; QRISK3; Sparse variable selection.

Graphical Abstract

Figure 1

Study design and flowchart. Cases are participants with a CVD event during follow-up including myocardial infarction and its sequelae, angina, non-haemorrhagic stroke, and transient ischaemic attack. Data were randomly split into three sex-stratified, non-overlapping sets: (i) variable selection data set (40%); (ii) training data set (30%), in which Cox models using selected variables were fitted; and (iii) hold-out test data set (30%), in which the predictive accuracy of these models was evaluated and compared with PCEs and QRISK3.

Figure 2

Stability selection LASSO. Selection proportions from LASSO stability selection calculated from (N = 1000) subsamples in (N = 48 689) men (left panel) and (N = 73 051) women (right panel). Explanatory variables considered include those contributing to PCE and QRISK3 scores (blue), genetic (brown), biochemical (green), and haematological (red) variables. Darker colours indicate stably selected variables (16 and 13 in men and women, respectively) as defined by variables with selection proportion above the calibrated threshold in selection proportion (vertical dark red dashed line).

Figure 3

CVD prediction in test data. ROC curves for logistic models predicting 10-year incident CVD in (N = 36 519) men (A) and (N = 54 792) women (B) in test data, where models use either recalibrated PCE (blue line), recalibrated QRISK3 (green line), or sex-specific stably selected variables (red line) in test data. We report the mean and 95% confidence intervals for the AUC.