Discussion
Relying on three commonly used 10-year CVD risk predictor algorithms and their variants,13 27 administered to a large cohort of ageing individuals residing in six SSA sites, this study found high levels of modifiable CVD risk factors with substantial variation and little agreement in predicted 10-year CVD risk across both the algorithms and the sites. This work furthermore highlights substantial differences in the levels of self-reported treatment of individuals identified to be at high risk, depending on the algorithm used to estimate risk—given the health service resource implications of this finding, it represents an important policy and clinical treatment consideration.
Variation in CVD risk by site and sex
While variation between the algorithms was observed in this study, general trends by algorithms persist, allowing us to compare estimated risk across sites. The results highlight varying levels of 10-year CVD risk by region, with lowest levels of estimated risk observed in the rural west African sites, higher levels in the urban Nairobi site and highest levels in the three South African sites. It is important to note that the rural South African sites had populations with higher levels of estimated risk than the urban Nairobi site across algorithms, which is perhaps surprising given the urban context of the Nairobi setting compared with the rural South African context. South Africa is well known to have a high and increasing burden of both CVD risk and morbidity and mortality.28 29 This study confirms these findings by demonstrating the highest level of CVD risk factors among the three South African sites, with more than one-quarter of each site’s population having at least three risk factors and more than one-third of Soweto’s population having at least three risk factors.
Our findings confirm the expected high level of CVD risk regardless of algorithm—a result of the epidemiological transition underway in Africa. Of concern, this study found nearly half of participants had at least two CVD risk factors and, depending on the algorithm used, the total 10-year risk of a future CVD event across all sites ranged from 2.6% to 6.5%, representing the population-level probability of experiencing a CVD event in the next 10 years. While our results likely reflect true inter-country variation in CVD risk, the interpretation of the level of risk at both the country and individual level remains obscured given the variation by algorithm as well as by the fact that the study population from each site was not nationally representative.
Little data from LMICs (especially from Africa) on 10-year CVD risk exist, making comparison of our results to previous studies difficult. However, the Research on Obesity and Diabetes among African Migrants (RODAM) Study has looked at CVD risk in Ghana, and when comparing our AWI-Gen Ghanaian findings to those of Boateng and colleagues,13 we found a lower proportion to be high risk using both the Framingham BMI and Framingham lipid algorithms (7.3% vs 19.4% and 1.8% vs 12.3%, respectively). Two possible reasons for these observed differences are the older ages included in the RODAM sample (40–70 years) versus the AWI-Gen sample (40–60 years), and the location of the study (both rural and urban in the RODAM study vs only rural Ghanaian participants in this study).30
Regardless of the algorithm employed, men were found to have higher levels of predicted 10-year risk than women across all sites. It is generally agreed that CVD affects women 7–10 years later than men31 and this is thought to be related to menopause, with endogenous oestrogen exposure delaying the development of atherosclerosis.32 Comparing CVD risk profiles of men to women 10 years older in high-income countries, one finds similar risk levels.33 This finding may differ in African populations as women from lower-income settings may undergo menopause earlier than women in higher-income settings.34 In the present study, the median (and mean) ages of study participants of both sexes were similar, likely resulting in the higher overall CVD risk seen in men. Yet as the African population continues to age, the expected increase in CVD risk between sexes may not be equal, with greater increases possible among women compared with men.35
Variation in CVD risk by algorithm
Using three commonly used CVD risk algorithms and their non lab-based variants, this study found substantial variation within sites, by algorithm. Comparing the lab and non-lab-based variants, there was generally better agreement, notably between the two Globorisk variants. The differences between algorithm variants are a likely result of the differing weight each risk factor contributes to the algorithm. Previous smaller African studies have examined differences between the two Framingham algorithms, and the pooled cohort equation algorithm in Ghanaian migrant and home populations13 and a single urban Kenyan site27 and our findings are similar with regards to the substantial variation of risk levels when using different algorithms. However, the current study goes beyond these earlier works by finding variation by algorithm across and within six diverse African sites. Previous studies also largely found similar trends in risk level—with Framingham office algorithms predicting higher CVD risk than laboratory algorithms.13 36 Previous cohort studies conducted outside of Africa have also shown the Framingham algorithms to overestimate coronary heart disease risk37 38 and require recalibration.39 In a large study from India, the Framingham algorithm was found to result in higher levels of risk compared with the Globorisk algorithm.40 Due to the fact that the WHO-CVD risk algorithm was only released in late 2019, to our knowledge, no study has sought to compare this new algorithm with other algorithms as was done in this study.
From this study, it is not possible to determine which algorithm most accurately estimates risk and there are likely to be a number of factors contributing to the variation in risk seen when using different algorithms. First, it is important to note that the calculated risk output definitions differ slightly by algorithm. For example, the Framingham algorithm output includes fatal and non-fatal CVD events including transient ischaemic attack and angina, while the Globorisk algorithms do not include transient ischaemic attack and angina.24 25 This could contribute to some of the increased risk levels seen when using the Framingham algorithm, but is unlikely to be the only factor. Another consideration likely contributing to the variation is the input variables. Framingham takes into account lipid levels, diabetes status and treatment for hypertension, while Globorisk does not, and WHO-CVD only includes diabetes status. Future analysis examining the role that the output and input factors described above play in predicting CVD risk is needed.
Each algorithm relies on coefficients derived from the analysis of population cohorts largely from the USA and Europe, with the Globorisk and WHO-CVD algorithms then re-calibrating the models using nationally representative health surveys or data from the GBD to derive more contextually relevant algorithms.14 15 24 However, data from GBD methods are challenged by studies being small, out of date, or reliant on modelling. Confirming the level of risk that each variable confers on an individual in a specific context would allow for the validation, or further recalibration, of these tools. It is possible that certain known CVD risk factors confer different levels of risk, depending on context and underlying genetics.41 42 However, there is a paucity of research on this topic on the African continent that hopefully studies like AWI-Gen, the CRSN Heidelberg Aging Study43 and the Health and Aging in Africa: A Longitudinal Study in an INDEPTH community12 44 can begin to address.
Levels of treatment by algorithm
In predicting population CVD treatment needs, it is important to choose a locally validated tool, given the substantial variation in risk levels that these results highlight. Using the recently released WHO-CVD risk prediction models, we see extremely low numbers of individuals with estimated high CVD risk across each study site. This has major cost and policy implications, as many treatment guidelines, such as WHO’s package for essential noncommunicable diseases in primary healthcare, which many LMICs use, take into account risk scores for evaluation and planned treatment.14 33 Based on the results of this study, the proportion of the population requiring treatment could vary by as much as 14% in urban South Africa, depending on the risk predictor algorithm used (Framingham BMI vs WHO-CVD (lab)). Given the recent release of the WHO-CVD algorithms, it will be important to examine the veracity of the CVD risk predictions generated by these algorithms over the coming years.
Strengths and limitations
This study has several strengths. First, the analysis was carried out using data from a large multisite, multicountry population-based sample of ageing Africans in diverse settings. The study employed rigorous, harmonised data collection methods in well-established research sites, ensuring that the input variables used in the algorithms were accurate, with the outcomes likely to represent the 10-year CVD risk in the studied populations. However, this study also has limitations. Treatment data from Soweto women are missing and having it would have allowed for a greater appreciation of treatment levels among urban women in a context of high CVD risk. Furthermore, the treatment analysis defined an individual as being under clinical management if they self-reported being on treatment for diabetes, hypertension or dyslipidaemia. We did not explore whether the individual was on the correct treatment for a specific condition. The intent of the analysis was to highlight the proportion of high-risk individuals, defined by the various 10-year CVD risk algorithms, that had also been identified by the healthcare system as requiring treatment. Related to this, it is possible that our estimates may have underestimated the proportion of high-risk individuals under clinical management as individuals who are on treatment may actually not be defined as ‘high risk’ due to the positive effects of their current treatment regimen. Second, many of the CVD risk factors rely on self-report and may be influenced by recall bias or social preference; however, the results of this work are largely aligned to previous findings from similar African studies.12 45 46 Finally, it is important to acknowledge that these population risk prediction levels are based on models, developed from studies largely carried out in high-income countries. Previous studies have suggested poor estimation of risk among minority groups.47 48 As such, employing these algorithms in LMICs, such as the six African sites of this study, may result in incorrect risk prediction. While the Globorisk and WHO-CVD algorithms attempt to take into account the likely contextual differences by re-calibrating their models using nationally representative data from diverse settings, more needs to be done to confirm the accuracy of the predicted risk in diverse populations, especially given the recent emergence of the WHO-CVD algorithm. The study populations in this research are not nationally representative and, as such, the results are unlikely to represent national levels of CVD risk. Only through the establishment of African cohorts and the careful follow-up and recording of CVD events, will these algorithms be validated, or adjusted, to account for the genetic, environmental, dietary and behavioural diversity found on the African continent. Context-specific CVD risk equations, or at least validation of existing tools, are needed as the African continent continues to experience rising levels of NCDs.