Suboptimal Down syndrome screening test interpretation

Ten years ago I wrote an Opinion for this journal entitled ‘Biochemical and ultrasound screening for Down’s syndrome: rivals or partners?' urging the combination of both modalities into a single late first-trimester test1. At that time the main obstacle to rapprochement was the method of test interpretation. Ultrasound nuchal translucency (NT) results were acted upon if the measurement, in mm, was increased above a fixed cut-off. In contrast, biochemical markers were standardized for gestation by expressing them as multiples of the gestation-specific median (MoMs), and the action cut-off was the risk of fetal Down syndrome given maternal age and the marker profile, estimated using a multivariate log Gaussian model. In the intervening years a strong partnership has been forged between the two modalities, principally through ultrasound converting its currency from mm to MoMs and risks. Statistical modelling predicts that routinely combining first-trimester maternal serum free beta-human chorionic gonadotropin (β-hCG) and pregnancy-associated plasma protein-A (PAPP-A) with NT yields, depending on gestation, an 80–87% detection rate for a 5% false-positive rate2. Prospective intervention studies have observed similar rates, although exact comparisons are not possible because of differences in the maternal age distribution, cut-off risk and gestation. Moreover, such intervention studies tend to overestimate detection by about 10% because almost one-half of affected pregnancies with risks below the cut-off miscarry and the fetus is not karyotyped. There have so far been eight studies (five cited in Reference 3, one later updated4 and four recent5-8) including a total of 77 000 women and 433 Down syndrome cases, with an overall 91% detection rate and 6.7% false-positive rate. But omens are increasingly seen suggesting that all is not sweetness and light in the relationship. There is controversy over whether NT interpretation should abandon MoMs in favor of delta units, arithmetic deviations from the gestation-specific median, and ipso facto abandon Gaussian risk calculation in favor of an empirical model9, 10. Delta units are also now being used for other ultrasound markers of Down syndrome such as ear length11 and maxillary bone length12. A paper by Leung et al. in the current issue of this journal on maxillary bone length uses normalized-delta (Z) units, arithmetic deviations divided by the SD, and diverges even further from biochemical screening practice by preferentially choosing menstrual dates over ultrasound biometry to define gestation13. These conflicts are particularly unfortunate when new ‘contingent’ Down syndrome screening strategies are being enthusiastically discussed, which would require even closer collaboration between the two modalities. With contingent screening, women at either high or borderline risk using markers tested at one stage in pregnancy are selected for further markers at later stages and the risk is revised accordingly. In one two-stage formulation, women with borderline risks based on first-trimester serum free β-hCG, PAPP-A and NT are selected to have four serum markers in the second trimester14. Modeling predicts that with just 15% of women in the borderline group the detection rate would increase to 94% for a 5% false-positive rate2. A three-stage modification has been proposed for areas where quality ultrasound resources are limited, splitting the initial stage into universal first-trimester serum free β-hCG and PAPP-A, and the selection of about one-third with the highest risk for ultrasound NT, followed by the same selection of those with borderline risks for second-trimester testing15. To improve detection further, newer ultrasound markers can be determined at the time of the NT scan: nasal bone, ductus venosus, and tricuspid regurgitation. However, these require much more experience than for NT measurement alone and a two-stage approach has been suggested where women with borderline risks based on first-trimester serum free β-hCG, PAPP-A and NT are selected for a more specialist scan16. A variant of this would be to move the NT into the second stage too and replace it in the first stage with an additional early serum marker. One candidate might be A disintegrin and metalloprotease 12 (ADAM12)17. When evaluating the optimal way to allow for gestation and calculate risks, three principles of screening need to be borne in mind. Firstly, screening as a public health measure aimed at reducing birth prevalence needs to maximize the detection rate for a given false-positive rate. This is achieved by minimizing the overlap between affected and unaffected pregnancies in the marker distributions. Secondly, screening as a means of individual risk assessment needs to ensure that the estimated risk is accurate; for example, the Down syndrome prevalence in women given 1 in 100–150 risks should be between 0.007% and 0.01%. It is possible to maximize detection whilst producing inaccurate risks for most women. Thirdly, in practice even the biggest screening center will not have a large enough throughput to be certain that they are achieving either their public health or individual risk assessment commitments. Hence, quality management is dependent on the adoption of, and adherence to, robust test interpretation methods. The level of most markers, ultrasound as well as biochemical, change with gestation and some method of gestational standardization is needed. There are really only two broad categories of method, multiplicative or additive, a ratio like MoM or a difference like delta units. And there is no a priori reason to choose one over the other; each application needs to be considered on its own merits. The first use of MoMs was in second-trimester maternal serum α-fetoprotein (AFP) screening for neural tube defects (NTDs). At mid-trimester, the level of AFP in maternal blood doubles every 5 weeks so some form of gestational age-standardization was essential. Moreover, it was readily seen that when MoMs were calculated from an ultrasound-based gestation, the normal range was much narrower than when menstrual dates were used. Thus there was less overlap between the distributions of NTD and unaffected pregnancies, hence a higher detection rate and a lower false-positive rate for the same 2.5 MoM cut-off. Furthermore, the biparietal diameter (BPD) is reduced in spina bifida, which leads to gestational underestimation, a systematic increase in MoMs, and a further detection increase. Down syndrome screening evolved as an extension of NTD screening and the use of MoMs was extended to the newer serum markers. A multivariate log Gaussian model was developed to estimate individual risks and their accuracy has been confirmed in nine studies3. Again, the benefit of ultrasound biometry over menstrual dating in reducing the overlap between the distribution of risks in affected and unaffected pregnancies was easily established. So large was the effect that a benefit could be seen even in women with certain dates, regular periods, and without pill withdrawal periods18. In Down syndrome, ultrasound biometric measures of gestation are identical to unaffected pregnancies except for the long bones, which are slightly reduced. Hence, when calculating MoMs, femur length (FL) and humerus length (HL) should not be used for dating since, in affected pregnancies, gestational underestimation will systematically increase AFP and unconjugated estriol, whilst reducing hCG or β-hCG and so underestimate risk. The best way to deal with FL and HL in this context is to express them in MoMs and, from a log Gaussian distribution model, calculate a combined risk together with the serum marker profile19. Some ultrasound authors try to simplify gestational standardization by using ratios other than MoMs such as BPD/FL, BPD/HL or BPD/nasal bone length, but since they change with gestation themselves there is no advantage. Some use the ratio between two markers, A and B, where A is elevated on average in affected pregnancies whilst B is reduced. Not only might this ratio change with gestation but it is also likely to be suboptimal even when it is constant. That is because one of the markers (say A) will be more discriminatory than the other, in so far as there will be less overlap between the distributions of Down syndrome and unaffected pregnancies. So, for example, the ratio will be two for a woman where A is twice normal and B is normal and one where A is normal and B is half normal, despite the first woman having the greater risk. Currently the main advocate of an additive rather than multiplicative method of gestational standardization is The Fetal Medicine Foundation (http://www.fetalmedicine.com) which has applied this to NT interpretation9. In Down syndrome pregnancies the average NT level declines between 11 and 13 weeks' gestation when expressed in MoMs: the meta-analysis estimates are 2.31, 2.10 and 1.91 MoM for each completed week2 whilst delta-NT appears to be reasonably constant. This does confer an advantage on delta-NT but constancy is not a necessary condition for optimal test interpretation. Over the same gestational period, the average serum free β-hCG and PAPP-A levels both increase when expressed in MoMs in Down syndrome pregnancies and no-one is suggesting that delta units be used. It is also claimed that the normal range in MoMs contracts with increasing gestation whilst the delta range is constant9, although that is not the experience of Leeds-affiliated centers. Among 13 530 pregnancies the normal 95th centile was no more constant for either method: 1.62, 1.45 and 1.44 MoMs compared with delta-NTs of 0.70, 0.65 and 0.73 mm at 11, 12 and 13 weeks, respectively. Whilst the dispute regarding NTs is usually couched in terms of how best to adjust for gestation, the real difference is in the approach to risk calculation. The distribution of NT MoMs is log Gaussian, at least over the 0.5–2.0 MoM range, whilst the distribution of delta-NT values is highly skewed and risks need to be calculated from an empirical model9. This considerably impairs the robustness of test interpretation. Empirical distributions will fit the original data set extremely well, including unique features of the population upon which it is based, and this limits applicability to other populations. In the case of NT measurement, the normal range varies from center to center20: in Leeds-affiliated centers the SD of log10 MoM is 0.09 compared with 0.12 in the population used to build the empirical model9. With a log Gaussian model this can be taken into account by altering the Down syndrome and unaffected SD, which are explicit parameters; in contrast, an empirical model is a one-fits-all approach. In practice, there is unlikely to be much difference when using NT MoMs compared with delta-NTs in the overlap of risks between Down syndrome and unaffected pregnancies. Most scans are carried out at the same gestation, 12 weeks, and both the Gaussian and empirical models are monotonic above the normal median, so that NT risks will be in a similar rank order when estimated by either method. However, the individual risks rather than their ranking will be altered by the method used and the relative accuracy of risk estimation by the two methods is unknown. Until the delta-NT approach has been shown to generate accurate risks in a large series other than that used to create the empirical model some doubt will remain. Leung et al.13 report an apparent correlation between ultrasound maxillary bone length, expressed in Z units, and concurrent serum PAPP-A MoMs among 607 women with reliable menstrual dates. The authors expected such a correlation because of a theoretical relationship between PAPP-A and bone growth. If confirmed, this would need to be taken into account when estimating Down syndrome risks; however, it could be a spurious observation due to confounding by gestational error. The two markers increase with gestation, so if a woman has an underestimated gestation both Z-maxillary bone length and PAPP-A MoM will be elevated and overestimation will have the opposite effect. Although less pronounced, gestational error still occurs in women with reliable menstrual dates but it can be reduced by ultrasound dating. This reduction in error can readily be seen by comparing Figures 2 and 4 in their paper, which shows that the within-gestation normal range of maxillary bone lengths is narrower using ultrasound dating compared to menstrual dates. No information is given about the shape of the Z-maxillary bone length distribution so it is not even possible to combine it with the serum marker profile for risk calculation. Since the discovery over 20 years ago that AFP was a marker of Down syndrome as well as NTDs, many additional biochemical and ultrasound markers have emerged and newer markers continue to be found. A prerequisite for these discoveries to be efficiently translated into improved prenatal diagnosis is the optimal interpretation of multimarker screening test results. Computer software can aid this process. It obviates the need to find simple rule-of-thumb methods of combining results and the risk calculation engine can easily handle gestational differences in model parameters. Software suppliers in the future will need to be flexible and provide both multiplicative and additive methods of standardizing for gestation, allowing the choice of the more appropriate and robust method for each marker.