Improving the yield of genetic testing in familial hypercholesterolaemia

This editorial refers to ‘Selection of individuals for genetic testing for familial hypercholesterolaemia: development and external validation of a prediction model for the presence of a mutation causing familial hypercholesterolaemia’, by J. Besseling et al., on page 565.

Introduction

Heterozygous familial hypercholesterolaemia (FH) is a common autosomal dominant disorder of LDL cholesterol (LDL-C) metabolism with an estimated prevalence of 1/500 to 1/200 in Caucasian populations.¹ It is caused by deleterious heterozygous mutations in genes affecting LDL receptor function (LDLR, APOB , or PCSK9).¹ As a consequence, LDL-C levels are elevated from birth, which leads to premature cardiovascular disease (CVD) in those affected.² Early interventions with lifestyle changes and statin treatment substantially attenuate this excess risk.³ Unfortunately, despite the WHO (World Health Organization) recommendation for large-scale screening almost two decades ago, FH is still severely underdiagnosed.¹ Genetic cascade screening of first-degree relatives is an effective way to improve the identification of affected subjects.⁴ However, cost represents a major barrier to its implementation in many countries.

The cost of genetic testing in an FH proband ranges from ∼€1600 in the UK⁶ to ∼€3100 in the USA.⁵ Cost-effectiveness analyses based on lipid and/or genetic screening approaches have estimated the incremental cost-effectiveness ratio for FH cascade screening to lie between ∼€3200 and ∼€460 000/life year gained.^5,6 This is thought to be determined by three key factors: (i) the cost-effectiveness of treatment once FH is identified; (ii) the prevalence of FH in the screened population; and (iii) the accuracy of the screening test.⁶ When using current diagnostic criteria to select patients for genetic testing, mutation detection rates are variable, but can be as low as ∼30%.^7,8 As such, increasing this yield has become a target for improving the cost-effectiveness of FH cascade screening.

Selecting individuals for genetic testing in FH: current approaches

Currently, the most widely used diagnostic criteria for FH include the UK Simon Broome FH Register (SB) criteria and the Dutch Lipid Clinic Network (DLCN) criteria.⁹ Variables considered include an LDL-C level threshold, medical and family history, and the presence of xanthomas and corneal arcus (Figure 1). The accuracy of these criteria in predicting the presence of an FH mutation varies substantially in different populations. The SB criteria provide two diagnostic classifications of FH: ‘definite’ and ‘possible’. A SB ‘definite’ classification is associated with mutation detection rates of 92% in Spanish subjects,¹⁰ 61–73% in Dutch and English subjects,^7,11 and 37% in Koreans.⁸ The specificity associated with this classification, however, is fairly uniform (82–93%), as is the negative predictive value (NPV) of not meeting the SB criteria for FH (52–88%).^7,8,10,11 The DLCN criteria provide three diagnostic classifications of FH: ‘definite, ‘probable’, and ‘possible’. A DLCN ‘definite’ classification based on clinical information is associated with mutation detection rates of >80% in English and Spanish subjects,^10,11 63% in Dutch subjects,⁷ and 35% in Koreans.⁸ Similarly to the SB criteria, the specificity of this classification in different populations is more uniform (77–88%), as is the NPV of receiving a ‘possible’ classification or not meeting the DLCN criteria for FH (60–94%).^7,8,10,11 Taken together, these findings show that FH diagnostic criteria perform consistently across populations in ‘ruling out’ an FH-causing mutation. Conversely, there is marked variability in how accurately current criteria predict the presence of an FH-causing mutation, especially in younger subjects and in populations outside the Western world. The poor predictive performance in these groups is, at least in part, attributable to the importance given to the presence of tendon xanthomas and family history, which are often either absent¹⁰ or difficult to ascertain,¹² and to levels of LDL-C that are lower than the set threshold.

Figure 1.

Variables assessed to determine the probability of genetically confirmed familial hypercholesterolaemia (FH) under current criteria vs. under the newly proposed model. Variables listed under ‘current criteria’ are a generalization of the Dutch Lipid Clinic Network (DLCN) and Simon Broome FH Register (SB) criteria. Variables listed are dichotomous unless otherwise indicated. CVD, cardiovascular disease, TC, total cholesterol; TG, triglyceride. ^‡Coded under four discrete categories: no CVD event, non-premature CVD event, CVD event at unknown age, premature CVD event.

Several modifications of the current diagnostic criteria have been proposed to improve the yield of genetic testing in FH, but have been met with mixed success.^10,11,13 The addition of a coronary calcium score and carotid intima-media thickness data to various clinical variables associated with FH did not substantially improve mutation prediction.¹¹ Moreover, these imaging modalities are not widely available, limiting their general applicability. A modified version of the DLCN maintained good predictive accuracy;¹³ however, it retains a requirement for information about tendon xanthomas and family history.

Selecting individuals for genetic testing in FH: a new model

In this issue of the journal, Besseling et al. present a novel approach to help identify patients carrying an FH-causing mutation.¹⁴ In contrast to the commonly used criteria, such as the SB and DLCN criteria, the newly developed model does not take into consideration information about tendon xanthomas or family history. Instead it uses routinely available lipid panel results and personal medical history.

To develop the model, the authors used a cohort comprised of all the subjects (>64 000) screened for an FH-causing mutation as part of the Dutch national FH screening programme between 1994 and 2014. The variables incorporated into their final model were chosen based on ease of availability and the strength of their correlation with a pathogenic FH mutation (Figure 1). In the development cohort, the model performed very well [area under the curve (AUC) 85.4%, 95% confidence interval (CI) 85.0–85.9] with good calibration (1.02). Of note, the model performed better at lower tertiles of age (AUC 18–40 years, 97.7% vs. 53–85 years, 92.9%) and higher tertiles of LDL-C (AUC 0.59–3.9 mmol/L, 52.8% vs. 7.06–25.14 mmol/L, 89.4%).

To validate the model, they chose a French Canadian cohort comprised of ∼3200 subjects that presented to an outpatient lipid clinic between 1993 and 2014. The prevalence of FH mutations in this cohort was comparable with that in the development cohort (French Canadian 44.8% vs. Dutch 40.8%); however, the subjects were older and with a higher prevalence of CVD and CV risk factors. In the validation cohort, the model performed better (95.4%, 95% CI 94.7–96.1%) with moderate calibration (1.06). However, the initial model underestimated the risk of a genetic mutation in the validation cohort and needed to be re-calibrated to account for the differences in the baseline risk for an FH mutation between the populations.

Unfortunately, the model developed by Besseling and colleagues was not directly compared with currently used criteria as part of their study; however, some preliminary observations can be made. The authors show that using a probability threshold of 0.7, their model is associated with a positive predictive value (PPV) of 86.1% (95% CI 85.9–86.4%), and a specificity of 94.5% (95% CI 94.4–94.6%),¹⁴ with even higher PPVs achievable at higher probability thresholds. Similarly, using a threshold of 0.3, the model gives rise to an NPV of 85.3%, while a threshold of 0.2 provides an NPV of 90%. These results are comparable with the best predictive values observed with current criteria. The superiority of the new model is also suggested by its discriminatory power. The model's AUC is ∼90% in both the validation and development cohorts.¹⁴ In contrast, AUCs of ∼65% and ∼70% have been associated with the SB and DLCN criteria, respectively.¹¹ Thus, these indirect comparisons suggest that the new model might outperform current diagnostic criteria in predicting the presence of an FH mutation.

Strengths and limitations of the new model

The major strength of the proposed model is its exclusive use of clinical and laboratory information routinely available in the medical records. Instead of relying on physical signs and family history, context for the identification of a pathogenic mutation is provided by a more detailed lipid panel (that, in addition to LDL-C, includes HDL-C and triglycerides) and medical history (that, in addition to premature CVD, includes history of hypertension, smoking, alcohol and statin use, and their interaction). This detailed information ensures that the model's output is appropriately adjusted for the potential confounding effects of non-FH dyslipidaemias or other CVD risk factors. Of note, in contrast to current criteria, this model performs best in young people. This can at least partly be explained by the inclusion of LDL-C as an age-adjusted continuous variable in the model and the exclusion of variables such as tendon xanthoma which are less prevalent in young people with FH.¹⁴

There are a few limitations that need to be addressed before this model can be implemented outside The Netherlands. First, the new model was not directly compared with current diagnostic criteria in the study. Secondly, further validation studies are needed. The validation cohort was comprised of patients attending a lipid clinic that were older, with more CV risk factors and a greater CVD burden than the development cohort. Initially the model substantially underestimated risk in this group, requiring re-calibration; this suggests that the original model is likely to require re-calibration to optimize risk prediction in non-Dutch populations. Finally, selection for genetic screening in the validation cohort was at the treating physician's discretion and not based on a clinical diagnosis of FH.¹⁴ This might have led to a selection bias, and contributed to the unexpectedly higher AUC in this group. The choice of a French Canadian founder population as a source of the validation cohort further limits the generalizability of the results from this cohort.

Conclusion and future directions

A model that accurately predicts the probability of having an FH-causing mutation is much needed to improve the cost-effectiveness of FH cascade screening. Besseling et al. have developed a model that relies on routinely available clinical information, performs well in young people, and has the capacity to achieve predictive values in excess of 90%. As such, this model has the potential to meet this important unmet need. However, further studies are much needed to establish a direct comparison with the currently used criteria and to validate this approach further in non-Dutch populations. Finally, given the strong correlation between clinical variables such as tendon xanthoma, family history of premature CVD, and the presence of FH-causing mutations and CVD,^10,11,15 it will be important to consider how the new model might be incorporated into current clinical practice alongside these important variables.

Funding

M.C. is supported by an NIH-NHLBI grant (HL059407).

Conflict of interest: none declared.