How far are we from accurate sex-specific risk prediction of cardiovascular disease? One size may not fit all

This editorial refers to ‘Sex inequalities in cardiovascular risk prediction’ by J. Elliott et al., https://doi.org/10.1093/cvr/cvae123. Since the term ‘risk factor’ was proposed for the first time in the Framingham study, multiple cardiovascular risk prediction models over the past decades have been developed to assess an individual’s future risk of developing cardiovascular disease (CVD), with the aim to implement certain interventions to those at elevated risk.1 Although sex-specific equations have been adopted in some of the predictive models, in the majority, the same risk variables are used for both males and females (albeit sometimes with different points for a given variable), and the intervention thresholds for males and females are often equal.2–4 Given the known sex differences in risk factors, metabolic physiology, and clinical outcomes associated with CVD, one size may not fit all. In this issue of the journal, Elliott et al.5 used UK Biobank data to evaluate the predictive performance of sparse sex-specific variables for CVD in comparison with the Pooled Cohort Equations (PCE) and QRISK3 CVD risk prediction models, with guideline specific thresholds. Their results showed that the model with sex-specific predictors had similar C-statistics in both males and females [i.e. 0.71 (0.70–0.72) for male; 0.72 (0.71–0.73) for female], equivalent or slightly higher than that of PCE [0.67 (0.66–0.68) for male; 0.69 (0.68–0.70) for female] and QRISK3 [0.70 (0.69–0.71) for male; 0.72 (0.71–0.73) for female]. Furthermore, the sensitivity using currently recommended risk thresholds in females was significantly lower than in males, and the use of a lower threshold was associated with increased sensitivity in females. This study highlights the development of a sex-specific predictive model with the selection of sparse variables, and the attempted use of sex-specific thresholds for more accurate prediction of CVD. Indeed, in addition to the commonly shared risk factors (including hypertension, diabetes, hyperlipidaemia, etc.), there are some female-specific risk factors such as pregnancy and a post-menopausal state, which may accentuate cardiovascular risk. The variation in risk factors for CVD between males and females have been demonstrated in previous studies.6 Female sex itself is an independent risk factor for CVD, and two-thirds of models had been explicitly developed for males and females separately, although they generally adopted the same factors.1 In the study, Elliott et al.5 used sex-specific predictors to develop the predictive model. Nevertheless, the predictive performance with the new model was not significantly improved compared with that of traditional models (all C-statistics around 0.7). Indeed, developing new models with improved performance is often challenging with complex designs and the use of additional variables without a marked improvement over older risk models, as shown by Elliott et al.5 In contrast, traditional models use some common but important risk factors—these offer the advantage of convenience and ease, whereas complex models with an increased number of variables, especially novel population-specific variables, may make the use of the models less convenient whilst also increasing the assessment cost, finally restricting its applicability. Therefore, an ideal cost-effective predictive model should first be practical especially when used for broad-based screening. The concept of sex-specific thresholds used for diagnosis and treatment of CVD is still worth exploring. One classic application is the sex-specific thresholds of troponin in patients with suspected acute coronary syndrome (ACS).7 In a study enrolling 48 282 patients suspected ACS, myocardial injury was defined as high-sensitivity cardiac troponin I concentrations > 99th centile of 16 ng/L in females and 34 ng/L in males patients; hence, use of this sex-specific threshold could identify five times more additional females than males with myocardial injury.7 In patients with atrial fibrillation (AF), females were historically at higher ischaemic stroke risk than males, leading to female sex being given an additional point as a risk modifier in the CHA2DS2-VASc score for stroke risk stratification.8 Nevertheless, more recent data show overall population stroke rates declining with marginal sex differences in the incidence of AF-related stroke, raising the possibility for female sex to be omitted as a factor when estimating ischaemic stroke risk and the need for oral anticoagulation therapy in patients with AF.8,9 Nevertheless, sex-specific thresholds could provide more accurate diagnostic or preventive value. This is particularly important for the prediction of CVD because inappropriate risk-based management may lead to overtreatment or undertreatment, whereas the development of a sex-specific predictive model with the use of sex-specific thresholds could improve predictive performance. As in the study by Elliott et al.5 the sensitivity of the predictive model in females was significantly enhanced when a lower risk threshold was used. Overall, the significantly lower true positive rate for females across all risk models in this study may have potentially skewed the performance, and further validation and assessment of a sparse selection of variables in different populations would be an important task for future studies. Developing a predictive model with greater accuracy is far beyond using sex-specific formulas and (or) thresholds. In the past decades, great efforts have been made to improve the accuracy of CVD prediction models, particularly through the addition of novel biomarkers such as biochemical measurements, imaging data, and genetic analysis, over and above traditional risk factors. However, the additional practical benefit provided by such markers has been shown to be limited and the cost-effectiveness is questionable, especially in the setting of population-wide screening tests. Also, statistical significance is not the same as clinical significance (and practical application). Risk factors also tend to cluster leading to ‘clinically complex’ risk phenotypes.10 In more recent years, machine learning has been shown to be a promising tool to improve the predictive performance and risk stratification for CVD, accounting for dynamic changes in risk factors over time.11 The latter is important given that risk is not static but changes with aging and incident comorbidities; however, traditional risk models investigate a risk factor at baseline and record event rates many years later. However, the generalization of machine learning to CVD risk prediction has a long way to go due to the heterogeneity in different population characteristics and the complexity of algorithms in machine learning. Currently, a universally predictive model applicable to all populations with a high performance is challenging. Derivation and validation of models requires a large number of participants from different regions, races, and socioeconomic backgrounds with abundant variables and continuous longitudinal observation spanning many years, accounting for dynamic changes in risk. This global, highly coordinated and collaborative work is expectedly extremely challenging. On the other hand, even if such a universally predictive model could be developed, this would likely be complex and may not be implemented conveniently as a routine screening tool. Questions remain on the performance of models using a smaller vs. a more exhaustive list of variables, as well as with sex-specific differences in variables included and thresholds used. However, the key sex differences in risk factors, biological and social determinants mean that we have further to go in the quest for a model that accurately and effectively takes that into account.