Cardiovascular risk estimation: can a risk prediction model derived in one country be used in another?

This editorial refers to ‘External validation and comparison of six cardiovascular risk prediction models in the Prospective Urban Rural Epidemiology (PURE)-Colombia study’, by J.P. Lopez-Lopez et al., https://doi.org/10.1093/eurjpc/zwae242. In this issue, Lopez-Lopez and colleagues1 explore an important issue in preventive cardiology, namely the extent to which the results of a cardiovascular risk prediction system derived in one country or region can be applied to a different population. The occurrence of atherosclerotic cardiovascular disease (ASCVD) in apparently healthy persons is usually the result of the combined effects of a number of risk factors. The clinical estimation of these combined effects is generally unreliable and for this reason guidelines on prevention recommend the use of a cardiovascular risk estimation system, six of which are assessed in the present paper-SCORE2, the AHA/ACC Pooled Cohort Equation, WHO, Globorisk Latin America, the Framingham Risk Score, and the non-laboratory INTERHEART Risk Score. Current systems for risk estimation generally use cohort studies starting at around age 40 to estimate the 10-year risk of a first ASCVD event. Regression techniques are used to define beta-coefficients, which are essentially multipliers used to express the independent effects of the risk factors under consideration. Core variables are age (exposure time rather than a risk factor per se), gender, smoking, blood lipids, and blood pressure. Other relevant variables that may be included are body weight, exercise, diabetes, ethnicity, and social deprivation. Some limitations of current approaches are considered later in this editorial. Strictly speaking, the results of a cohort-based risk estimation system apply only to the population from which it was derived, or to populations with closely similar characteristics. But if they cannot be used more widely, the whole exercise is pointless, which is why the current publication is of relevance, not just to Columbia but more generally. This issue had to be addressed in developing and re-calibrating the SCORE2 risk model for four different risk regions of Europe.2 Conventional approaches to recalibration require up-to date risk factor, mortality and non-fatal event data.3 Of these, non-fatal events are most challenging in view of variations in data quality and ascertainment methods. These issues may have posed challenges for both the Globorisk and WHO risk models. The current paper describes an interesting contrasting approach to recalibration, as will be seen. Lopez-Lopez and colleagues reviewed the current literature on the evaluation of existing risk models for use in Columbia and found appreciable limitations. They therefore examined the ability of the six cardiovascular risk prediction models to estimate risk in 3802 subjects without ASCVD at baseline of the 7552 subjects in Prospective Urban Rural Epidemiology (PURE)-Colombia study. Most exclusion related to missing values, especially lipids. Those with diabetes or who were taking statins were excluded, but a further examination indicated the inclusion of these would have had a very small influence on the results. Subjects with a systolic blood pressure of >200 or <90 mmHg were excluded, as were those with an LDL cholesterol of >342 or <115 mg/dL (9 and 3 mmol/L, respectively). The performance of the models was assessed by means of conventional methods—discrimination (the ability of the model to correctly classify a positive or negative event, in this case using the C-statistic as a summary of the integrated area under the Receiver Operating curve) and calibration (the degree of similarity between the observed and predicted results). While widely used, the C-statistic has limitations. It is essentially an expression of the tradeoff between sensitivity and specificity; management decisions will be obvious at the extremes of risk whereas guidance is needed close to thresholds at which an intervention is being considered. One possible solution is to use the Net Reclassification Index.4 This was not employed in the current paper and indeed has been criticized.5 All six prediction models were reported to show similar discrimination. With regards to calibration, all models over-estimated risk, least with SCORE2 and most in Globorisk-LAC, WHO and the AHA/ACC PCE. In view of the generally similar and acceptable discrimination, it was felt appropriate to recalibrate each of the prediction models using a contrasting technique to that used in SCORE2, namely an Integrated Calibration Index,6 resulting in individual correction factors for each model. Recalibration factors varied from 0.75 (SCORE2, Women) to 0.27 (Framingham Risk Score, Men). A reasonable question is whether the use of risk estimation models, in Columbia or elsewhere, logical though it may be, results in improved patient outcomes. This would require a randomized control trial comparing usual care with risk-prediction driven care and it is unlikely that such a trial will be undertaken-the last trial of a total risk approach to care was MRFIT7 20 years ago and undertaken at a time that ASCVD mortality was falling and few effective drug treatments were available. What of the future? Whatever risk estimation system may be adopted in Columbia, there is a need to incorporate measures of social deprivation, inflammation, and probably LPa. Imaging such as CAC can improve risk prediction as a diagnostic test for asymptomatic disease rather than a risk factor per se. Its use may stimulate more intensive risk factor management. Present methods may have gone as far as they can. A move from 10-year risk to a cumulative exposure time model integrated with Mendelian Randomization and utilizing artificial intelligence to explore interaction effects may be the future, with the hope of allowing more precise personalized risk estimates from earlier in life than is presently possible.8,9

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.