Noninvasive Fetal Sex Determination Using Cell-Free Fetal DNA:  A Systematic Review and Meta-analysis FREE

Noninvasive prenatal determination of fetal sex could provide an important alternative to invasive cytogenetic determination, which is currently the gold standard for ambiguous genitalia, X-linked conditions, and single-gene disorders such as congenital adrenal hyperplasia. Chorionic villus sampling and amniocentesis have small but measurable rates of procedure-related pregnancy loss.^1- 3 The availability of a reliable noninvasive alternative to determine fetal sex would reduce unintended fetal losses and would presumably be welcomed by pregnant women carrying fetuses at risk for disorders. A much broader potential application for fetal sex detection is family balancing, which poses ethical concerns.

Fetal sex determination can be performed by sonography at as early as 11 weeks' gestation, although not reliably. Test performance across published studies varies significantly. According to a review by Odeh et al,⁴ fetal sex cannot be determined by ultrasound examination in 7.5% to 50% of pregnancies at 11 weeks' gestation, and this decreases to 3% to 24% at 13 weeks. When reported, the sex determination is incorrect as often as 40% of the time at 11 weeks, although by 13 weeks, accuracy (when reported) is close to 100%. A subsequent prospective study reported that fetal sex could be determined by ultrasound 90% of the time, with 86% accuracy from 11 to 14 weeks' gestation.⁵

The presence of cell-free circulating Y chromosome DNA sequences in the plasma of pregnant women was first described in 1997.⁶ Since that report, many groups worldwide have validated the initial finding that Y chromosome sequences can be amplified and used to identify male fetuses. The research has been extended using a variety of methodologies, sex-specific markers, and sample types across all gestational ages. In countries such as the Netherlands, the United Kingdom, France, and Spain, this testing has already transitioned to routine clinical care despite the absence of a formal assessment of its performance.⁷ More recently, companies have begun offering this technology directly to the consumer over the Internet.⁸ The tests are marketed for nonmedical use to curious parents-to-be with promises in some cases of accuracy as high as 95% to 99% at as early as 5 to 7 weeks' gestation.^8- 9

To document the performance of noninvasive fetal sex determination using cell-free fetal DNA, identify the variables that affect performance, and identify sources of heterogeneity between studies, we performed a systematic review and meta-analysis. We sought to determine the analytic validity of cell-free fetal DNA testing, which describes the test's ability to detect Y chromosome sequences within maternal samples, as well as the clinical validity of the test, as indicated by its ability to correctly identify fetal sex. Primarily, the studies focused on the overarching question, “How reliably can fetal sex be predicted by sex-specific markers in cell-free fetal DNA from a maternal blood or urine sample?”

We searched the PubMed database for English-language studies containing primary data involving humans. The search included articles published between January 1, 1997, and April 17, 2011. The first report of the presence of fetal DNA in maternal blood was published in 1997.⁶ The following search terms were used: cell-free fetal DNA, fetal sex DNA, prenatal sex DNA, fetal sex detection, cell-free prenatal DNA, non-invasive fetal sex, and non-invasive prenatal. The alternate spelling foetal was also used. Reference lists from review articles were searched for additional relevant publications.

Two authors (S.A.D., D.W.B.) read the abstracts of each study. Review articles and others that did not contain primary data were excluded. Each abstract was screened for specific key words representing inclusion criteria: DNA, fetal sex detection, Y chromosome sequences, pregnancy, blood, plasma, serum, amelogenin, and quantitative PCR. Amelogenin is a protein encoded by a gene present on both sex chromosomes but with a 6–base pair deletion on the X chromosome version (the genes are AMELX and AMELY on the X and Y chromosomes, respectively). Because the focus of our study was on blood, urine, DNA, and pregnancy, each identified abstract was then screened for the following key words representing exclusion criteria: RNA, amniotic fluid, peritoneal fluid, Rhesus D, single gene disorder, formaldehyde, beta thalassemia, beta globin, fetal cells, leptin, post-partum, and spontaneous abortion. The covariates chosen were publication year, sample type, DNA amplification methodology, Y chromosome sequence, and gestational age.

The primary reviewer (D.W.B.) performed the preliminary extraction of data from each selected study using a standard form. The review form was designed to capture primary data, including the number of samples from women with singleton male and female fetuses, the number of each postulated to be male, the number of indeterminate cases, and the cause of false-negative or false-positive results. Additional data included sample type, methodology (sample handling, polymerase chain reaction [PCR] technique, amplified gene sequence, housekeeping gene[s], type[s] of control[s] used, the number of replicates, and the referent test [ie, the method used to confirm fetal sex]), and any other clinical factors described in the study, such as preeclampsia or preterm labor. The primary reviewer then made a judgment about whether the study contained data useful for the subsequent analysis of clinical validity. Using the same form, a second reviewer (S.A.D.) also extracted the data to be used in the meta-analysis. Interreviewer discrepancies were resolved by discussion.

The primary reviewer assessed study quality with a standardized form. The reviewer rated aspects of study design across 23 relevant criteria that have implications for analytic validity and clinical validity using a published methodology for grading evidence from genetic tests.¹⁰ We then applied numeric values to the grades (unsure = 0, inadequate = 0, adequate = 1, and convincing = 2), for a potential total score of 46. We divided the total possible score into tertiles: 0 through 15 (low quality), 16 through 30 (average quality), and 31 through 46 (high quality).

The data were analyzed using Meta-DiSc version 1.4 (available at http://www.hrc.es/investigacion/metadisc_en.htm).¹¹ We calculated the sensitivity (proportion of male fetuses with a positive test result) and specificity (proportion of female fetuses with a negative test result) for each included data set. Indeterminate test results were considered to be false-positive or false-negative depending on the true fetal sex. The threshold for calling a result indeterminate may differ between studies. However, we did not account for this variable, because indeterminate results were uncommon and the threshold used for making a determination was not often reported.

Given the relationship between sensitivity and specificity introduced by a potential threshold effect, the diagnostic odds ratio (OR) was used as the primary method of calculating the summary test performance. We decided a priori that we would not rely solely on summary sensitivities and specificities, because groups are likely to have used different cutoffs to identify a positive test result (threshold effect). To test for this possibility, we plotted study-specific sensitivity and specificity on a summary receiver operator characteristic (ROC) curve and looked for a curvilinear shape that is characteristic of the threshold effect. The curve was fit to the data using the Moses-Shapiro-Littenberg model.¹² Any difference in the threshold used to call a result indeterminate is not accounted for by the summary ROC curve. The Spearman correlation coefficient was used to assess the existence of an inverse relationship. Area under curve (AUC) values and asymmetric confidence intervals (CIs) were calculated as previously described.¹³

To deal with zero observations in 2 × 2 contingency tables, ½ was added to each cell, reducing performance in the small studies. Six outliers outside the 99% CI of the summary ROC curve were identified and removed from further analyses. In 3 of the outlier data sets, urine was used as the sample type, which is unreliable for detecting cell-free fetal DNA.^14- 15 Two data sets had high rates of indeterminate results (24% and 7.4%),^16- 17 and this translated into high rates of false-positive and false-negative results. The remaining data set reported a number of false-positive and false-negative results; in this study, whole-blood samples were collected between 7 and 11 weeks' gestation, and nested PCR used for the amelogenin-encoding gene was performed.¹⁸ We removed the outlier data sets so they would not have undue influence on the summary ROC curve. The 99% CI was chosen to ensure that only the few most aberrant results would be excluded from the analysis. To reduce the potential for publication bias, we set an a priori inclusion criterion of at least 10 male samples and 10 female samples per study.

Heterogeneity was evaluated using I², the proportion of heterogeneity not explained by random chance. I² is calculated from the H statistic, which is the square root of the commonly used χ² statistic (also known as Q test of heterogeneity).¹⁹I² is the interpretation of the χ² value, accounting for the df.

Positive predictive value (PPV) (true positives/[true positives + false positives]) and negative predictive value (NPV) (true negatives/[true negatives + false negatives]) were calculated based on the observed rate for each test. The Q* index, derived from the fitted curve, is the point at which sensitivity equals specificity and was calculated for each test. Because of the threshold effect, we wanted a summary measure of test performance based on the fitted ROC curve.^20- 21 Because the test for fetal sex determination can have high performance, the Q* index usually falls in a clinically relevant region (high sensitivity and specificity). The alternative would be to choose a fixed sensitivity (or specificity), but the choice of appropriate values would be dependent on the setting in which the test is used. Independently pooling sensitivity and specificity across multiple studies may be misleading, because studies used varying cutoff levels to define a positive test result (threshold effect). Fitting the paired results to a summary ROC curve minimizes this issue and allows for more reliable performance estimates. The Q* index allows for a single number to summarize test performance, facilitating comparisons. Level of significance was set at P < .05. Testing was 2-sided.

To determine whether any single data set was incurring undue weight in the analysis, we systematically removed 1 data set at a time and computed Q* index for the remaining group. This was conducted for each data set.

Prior to analysis, the following potential covariates were chosen: publication date (2003 or earlier, 2004 or later), based on an observed change in study size (Figure 1); sample type (plasma, serum, whole blood, or urine); amplification technique (conventional PCR or real-time quantitative PCR [RTQ-PCR]); Y chromosome sequence (SRY [GenBank NG_011751.1], DYS14 [GenBank NG_027958.1], DYS1/DAZ [GenBank NG_004755.2], DYZ3 [GenBank NG_004755.2], and AMELY [GenBank NG_008011.1]); and gestational age in completed weeks. We chose the following gestational age ranges: early gestation (<7 weeks), 2 clinically relevant ranges (prechorionic villus sampling [7-12 weeks] and preamniocentesis [13-20 weeks]), and late gestation (>20 weeks). Each data set was coded for these covariates. The diagnostic OR was compared between covariate groups using a random-effects model.

To investigate whether the statistically significant univariate effects were independent, we performed a meta-regression using the Meta-Disc software.¹¹ Meta-Disc uses the Moses-Shapiro-Littenberg method.¹² Studies with missing data were excluded from the meta-regression.

The primary literature search yielded 330 published studies. An additional 529 were identified by searching the reference lists of review articles, for a total of 859 studies. We excluded 713 based on abstract content alone. The remaining 146 studies were read in their entirety (eTable 1). A total of 74 studies^{6- 7,14- 18,22- 88} were suitable for inclusion in the analyses (Figure 2). Interreviewer agreement was high. There were 1504 data points collected from the 74 studies, and only 15 discrepancies (1.0%) occurred. All were resolved through discussion.

Sufficient information was not available to calculate analytic validity. Although a handful of studies did address this issue, most of the studies focused on clinical validity (ie, how reliably fetal sex could be predicted). Imperfect analytical validity will be apparent when assessing clinical validity, because incorrect fetal sex assignments will nearly always be due to analytic error. Therefore, a separate assessment of analytic validity is not necessary in this setting.

Of the 74 studies, 5 were excluded for having insufficient data: 4 had data on sensitivity only (male fetuses only)^49,73,85,87 and 1 had data on specificity only (female fetuses only).⁶⁶ Two additional studies^74,82 were excluded because the design was of poor quality; analysis was not blinded in one, and no controls were used in the other. This left 67 studies. Ten small studies^{33- 35,37,56- 57,60- 61,67,77} were removed because they included fewer than 10 male fetuses and 10 female fetuses. These also tended to be pilot trials or studies that included only male fetuses or only female fetuses. This left 57 studies for the analyses. Fifteen^{6,14- 17,23,25,27,55,59,62,70,75,78,88} of these 57 studies could be stratified into multiple data sets. For example, a study that tested the performance of 2 different Y chromosome sequences on the same samples was divided into 2 data sets. Therefore, from the 57 selected studies used in the meta-analysis, we extracted 80 data sets (eTable 2). We treated all data sets as independent for the univariate analyses. Among these data sets were discrete samples from pregnant women bearing 3524 singleton male and 3017 singleton female fetuses.

The sensitivity and specificity of fetal sex determination in each of the 80 data sets was calculated and plotted on a summary ROC curve (Figure 3). The Q* index is represented by the off-diagonal line. Six outlier data sets (outside the 99% CI) were removed,^14- 18 3 of which were derived from studies using maternal urine as the sample type, leaving 74 data sets. There were 7 data sets for which urine samples were tested, including the 3 outliers; 2 others that had been excluded for having fewer than 10 male samples, fewer than 10 female samples, or both^35,60; 1 that had been removed previously for having data on sensitivity only⁴⁹; and 1 additional data set.⁷⁹ We analyzed urine separately; thus, the additional urine data set was removed,⁷⁹ leaving 73 data sets reporting test results from blood components for analysis. Five data sets^16,25,62 (described below) were removed for using a method other than conventional PCR or RTQ-PCR, leaving 68 data sets from 52 studies for the univariate analyses (eTable 2).

The summary ROC curve (Figure 3) illustrates the significant between-study heterogeneity in test performance (for overall diagnostic OR, χ² = 108, P < .001, and I² = 38%). Sensitivities ranged from 65% to 100%, and specificities from 73% to 100%. The AUC was 0.993 (95% CI, 0.989-0.995). Both χ² and I² showed more heterogeneity than expected by chance alone.

We examined whether any one data set had undue influence on the analysis by recomputing Q* index repeatedly, excluding a different data set on each pass. The data set with the largest influence increased Q* index from 96.84% to 97.06% (+0.22%)⁵⁵; the next largest reduced Q* index to 96.63% (−0.21%).⁷⁰ This indicates that no one study had a large influence on overall test performance.

We assessed the 68 data sets included in the univariate analyses for quality of study design. The studies were stratified by overall score into 3 groups: 0 through 15 (poor quality [n = 43 data sets]), 16 through 30 (average quality [n = 22 data sets]), and 31 through 46 (high quality [n = 3 data sets]). There was no significant difference between the performance of data sets that scored 0 through 15 (diagnostic OR, 903; PPV, 98.9%; NPV, 96.7%), those that scored 16 through 30 (diagnostic OR, 863; PPV, 99.2%; NPV, 93.5%), and those that scored 31 through 46 (diagnostic OR, 657; PPV, 94.9%; NPV, 87.0%) (P = .96 for comparison of diagnostic ORs) (Table 1).

To examine potential sources of heterogeneity not accounted for by random chance, the 68 data sets were stratified by the following 5 covariates: year of publication, sample type, amplification technique, Y chromosome sequence, and gestational age. The sensitivity, specificity, diagnostic OR, PPV, NPV, I², and Q* index were stratified by each potential covariate (Table 1). The diagnostic OR and the Q* index were used because they take into account both sensitivity and specificity, which on their own can be less informative as summary statistics because of the threshold effect. The diagnostic OR can be interpreted as the increase in the odds that a fetus determined by the test to be male is actually male.¹¹

Of the 57 studies (80 data sets) plotted on the summary ROC curve, only 7 reported indeterminate results.^{16- 17,28,42- 43,80,84} The average rate of indeterminate results across these 7 studies was 6.8%. The threshold chosen to make a determination of fetal sex will affect the test failure rate; more stringent thresholds will increase the likelihood of indeterminate results. Two of the outlying studies excluded had high failure rates (7.4% and 24%) attributable to indeterminate results.^16- 17

Studies were grouped based on their publication date (2003 or earlier, 2004 or later). The year 2003 corresponded to a reduction in the number of studies published that year, as well as a general change in study sizes and methodologies. With one exception, most of the large studies (n  150) were conducted in 2004 or later (Figure 1). In addition, the year of publication may be an indirect measure of improving technology, methodology, or both. Twenty-one of the data sets were published before 2004, and 47 were published in 2004 or later. There was no significant difference between the performance of studies published in 2004 and later (diagnostic OR, 1055; PPV, 98.7%; NPV, 95.5%) compared with those published in 2003 and earlier (diagnostic OR, 532; PPV, 99.1%; NPV, 92.4%) (P = .15 for comparison of diagnostic ORs) (Table 1).

In studies for which the samples were derived from maternal blood, for 11 data sets, maternal serum was tested; for 55 data sets, plasma was tested; and for 2 data sets, whole blood was tested. For 7 data sets, urine samples were tested. Plasma (diagnostic OR, 997; PPV, 98.8%; NPV, 95.3%) and serum (diagnostic OR, 776; PPV, 98.9%; NPV, 94.6%) performed similarly across studies (P = .66 for comparison of diagnostic ORs). Tests using whole blood were only available for 2 data sets.^22,68 The CIs for summary sensitivity (85.7% [95% CI, 78.1%-91.5%]) and specificity (92.4% [95% CI, 83.2%-97.5%]) for whole blood were too large to allow drawing any conclusions on the true test performance, so we excluded the 2 data sets prior to performing the meta-regression (Table 1).

Of the 7 data sets for which maternal urine was used, 2 were excluded for having fewer than 10 male samples, fewer than 10 female samples, or both, and 1 data set was excluded for testing males only (ie, only sensitivity data were available).^35,49,60 The remaining 4 data sets^14- 15,79 showed poor and inconsistent performance (sensitivity: 6.3%, 32.3%, 38.2%, and 95.1%; specificity: 98.2%, 98.2%, 96%, and 87.9%; and PPV, 90.7% and NPV, 52.4% [across the 4 data sets]). The high specificity is likely an artifact; if no fetal DNA can be amplified, the test will not produce false-positive results.

Conventional PCR amplification was reported used in 25 data sets and RTQ-PCR in 43 data sets. Four studies (5 data sets) used alternate methods: probe microplate hybridization,²⁵ matrix-assisted laser desorption/ionization time-of-flight mass spectrometry,^62,84 and pyrophosphorolysis-activated polymerization.¹⁶ We excluded these 5 data sets because of insufficient data for analysis. RTQ-PCR (diagnostic OR, 1284; PPV, 99.1%; NPV, 95.5%; AUC, 0.996 [95% CI, 0.993-0.998]) performed significantly better than conventional PCR (diagnostic OR, 424; PPV, 97.8%; NPV, 92.6%; AUC, 0.988 [95% CI, 0.979-0.993]) for predicting fetal sex (P = .02 for comparison of AUCs; P = .01 for comparison of diagnostic ORs) (Table 1).

Five different Y chromosome sequences were used: SRY (n = 31), DYS14 (n = 21), DYS1 /DAZ (n = 9), DYZ3 (n = 4), and AMELY (n = 3). Among the 146 studies reviewed in full, only 5 used a Y chromosome sequence other than the 5 listed above: ZFX, ZFY⁷⁴; short tandem repeat DYS390⁸⁹; Y sequence primer pairs Y1.5/Y1.6⁴⁶; CYorf15B, NLGN4Y, XKRY⁸²; and DBY, TTTY.2.⁸⁴ There was no significant difference between the test performance of SRY (diagnostic OR, 996; PPV, 99.3%; NPV, 94.5%), DYS14 (diagnostic OR, 701; PPV, 97.9%; NPV, 94.6%), and DYS1/DAZ (diagnostic OR, 1387; PPV, 99.2%; NPV, 97.5%) (P = .61 for comparison of diagnostic ORs). The other 2 genes, DYZ3 and AMELY, performed well (DYZ3 [n = 4]: diagnostic OR, 646; PPV, 98.4%, NPV, 95.0%; AMELY [n = 3]: diagnostic OR, 802; PPV, 98.3%, NPV, 95.2%). However, the CIs were too large to be informative (DYZ3: sensitivity, 95.4% [95% CI, 90.2%-98.3%]; specificity, 98.3% [95% CI, 93.9%-99.8%]; AMELY: sensitivity, 95.9% [95% CI, 90.8%-98.7%]; specificity, 98.0% [95% CI, 93.0%-99.8%]). We removed these 7 data sets^{27,53,59,68,78} prior to meta-regression (Table 1).

Gestational age was divided into 4 different ranges: early gestation (<7 weeks); 2 clinically relevant ranges (prechorionic villus sampling [7-12 weeks] and preamniocentesis [13-20 weeks]); and late gestation (>20 weeks). In the 30 studies included, the researchers had either tested pregnant women within a gestational age range that fit our categories, or they provided the individual data to be properly categorized. A total of 49 data sets were analyzed: 4 from less than 7 weeks, 21 from 7 through 12 weeks, 16 from 13 through 20 weeks, and 8 from more than 20 weeks.

The performance of the test increased with gestational age. The 7 data sets that reported results prior to 7 weeks' gestation had poor performance. The 4 studies included in our univariate analysis^{63,71- 72,88} (the other 3 were excluded for insufficient sample numbers^36,65,69) showed low sensitivity: 74.5% (53.9%, 72.7%, 72.7%, and 88.9%, respectively). The specificity was 99.1%, which is probably artificially high as a result of no fetal DNA amplification at such early gestation (Table 1). For gestational age less than 7 weeks (n = 4), the AUC was 0.989 (95% CI, 0.965-0.998). In the late first trimester (7 through 12 weeks, n = 21) (diagnostic OR, 846; PPV, 98.9%; NPV, 95.1%; AUC, 0.994 [95% CI, 0.987-0.997]) and early second trimester (13 through 20 weeks, n = 16) (diagnostic OR, 604; PPV, 99.4%; NPV, 93.3%; AUC, 0.992 [95% CI, 0.983-0.996]), the tests performed equally well (P = .57 for comparison of diagnostic ORs). The best test performance was demonstrated in women at 20 weeks' gestation and later (n = 8) (sensitivity, 99%; specificity, 99.6%; diagnostic OR, 3196; PPV, 99.6%; NPV, 98.9%; AUC, 0.998 [95% CI, 0.990-0.999]) (P = .02 for comparison of diagnostic ORs across 4 gestational age ranges) (Table 1).

Meta-regression was performed on the 46 data sets that had available data for publication year, sample type, DNA amplication methodology, gene, and gestational age (eTable 2). This represents 60% of the original 6541 pregnancies identified in the literature search. We accounted for duplicate data sets, in which the same samples were used for different tests, by stratification. Based on a small number of studies and limited predictiveness of gestational age and method, we chose not to include interaction terms. Only amplification technique and gestational age were found to be significant predictors of test performance. Year of publication (diagnostic OR, 0.87 [95% CI, 0.25-3.01; P = .83]), sample type (diagnostic OR, 1.19 [95% CI, 0.27-5.15; P = .82]), and gene (diagnostic OR, 0.81 [95% CI, 0.34-1.92; P = .62]) were not. Based on these results, we restricted the multivariate model to amplification technique (diagnostic OR, 3.39 [95% CI, 1.08-10.66; P = .04]) and gestational age (diagnostic OR, 6.06 [95% CI, 2.24-16.38; P < .001]). We then modeled the diagnostic OR for various combinations (Table 2). The highest performance is achieved after 20 weeks' gestation using RTQ-PCR, with a diagnostic OR of 5388.

This systematic review identified 146 publications that could be used to determine the clinical validity of noninvasive prenatal sex determination using cell-free fetal DNA in maternal blood and urine. Possible limitations of our search and potential sources of bias include the use of only 1 medical database, which may result in omission of studies indexed elsewhere^90- 91; exclusion of unpublished reports; lack of contact with experts for additional studies not found in the electronic search; exclusion of publications not written in English; and a limited exploration of potential publication bias. In addition, the 2 reviewers were not blinded to the article citation during review.

Despite the variability between studies, the overall sensitivity (95.4%) and specificity (98.6%) is high (Table 1). Testing performed prior to 7 weeks' gestation using blood, and all tests using urine, are unreliable.

Univariate analyses showed that amplification technique and gestational age had the strongest effects on test performance, that RTQ-PCR outperformed conventional PCR, and that performance of the test was much lower prior to 7 weeks. Regression analysis and modeling results found that these 2 variables in testing are responsible for some of the heterogeneity observed across studies. The best performance is predicted with use of RTQ-PCR after 20 weeks' gestation. The performance difference between amplification methods is likely attributable to the greater dynamic range of detection of RTQ-PCR vs most traditional PCR methods. The improved performance with later gestation is likely attributable to the increased concentration of cell-free fetal DNA within maternal blood as the fetus and placenta develop. This would explain the poor performance of the test prior to 7 weeks' gestation and the near-perfect performance in the third trimester. Date of publication, sample type (serum vs plasma), and gene target did not have significant effects on test performance.

One variable unaccounted for is interlaboratory variability in the criteria for positive and negative determinations. This can be seen in the discordant reporting of indeterminate results. Within our review of the studies, we noted how the researchers handled indeterminate results. This included reading the study for any hints that samples may have been removed from the reported results. We reviewed the studies with this in mind and captured all indeterminate results, whether explicitly stated or made apparent through the data. Apart from the 7 studies with indeterminate results, none of the others reported or made reference to indeterminate results or otherwise indicated samples for which results were not reported. Because the majority of studies do not report the threshold used to make a determination or to declare a result indeterminate, it is difficult to discern if a determination is made for every sample or if indeterminate results might not be reported. If the latter is true, then our understanding of the true failure rate is limited. This will be important to know in the event that this testing transitions to clinical care in the United States.

One study that reported a 5.4% failure rate used a strict threshold to make a positive determination: amplification of both DYS14 and SRY.⁸⁰ For tests in which there was no amplification, paternal and maternal biallelic polymorphisms were used to validate the presence of fetal DNA. If fetal DNA could not be validated, the sample was called indeterminate. This example illustrates how the likelihood of obtaining indeterminate results increases as the stringency for making a positive determination increases. The opposite of this is evidenced through poorly performing tests, such as tests conducted before 7 weeks' gestation. For these tests, sensitivity was low (74.5%) but specificity was high (99.1%). It is probable that the concentration of cell-free fetal DNA in some samples was undetectable, but this is not reflected in test specificity. Use of this technology for clinical diagnosis should require stringent criteria that provide validation of negative results. Before calling a sample female, the laboratory should show that fetal DNA is present and can be amplified.

Clinical laboratories also should consider the reason for the test when setting the threshold for positive determinations. For X-linked disease, it is more important to eliminate false-negative results (eg, calling a male fetus female), because all male-bearing pregnancies will need follow-up invasive testing. Therefore, the threshold should be set to favor sensitivity at the deficit of specificity. In the case of congenital adrenal hyperplasia, which affects female genitalia in utero, specificity should be favored.

Currently, the standard non-invasive method of determining fetal sex is by ultrasound examination. Typically, fetal sonographic studies are performed as part of routine maternal care with the intention of identifying fetal anomalies. Fetal sex is often provided as part of the analysis. Advantages of ultrasound examination are that it is noninvasive, widely available, and accurate after approximately 13 weeks' gestation. However, some studies have shown that diagnosis of fetal sex is not always possible, even at 13 weeks.⁴ One clear disadvantage is that ultrasound is unreliable before 11 weeks' gestation. For women who are carriers of X-linked diseases and require invasive testing, sex determination earlier than 13 weeks is critical to reduce invasive testing. Our review of the data regarding amplification of Y chromosome markers from maternal blood indicates that this can be performed reliably between 7 and 12 weeks' gestation, earlier than with sonography. A disadvantage of maternal blood testing is the need to verify the presence of fetal DNA to validate determinations of female sex. Other potential disadvantages include that DNA testing is not available at point of care, not presently approved by the Clinical Laboratory Improvement Amendments, and not currently reimbursed by insurers.

Many of the studies included here were relatively small. A large, prospective, randomized, blinded clinical trial would be beneficial to help validate test performance under highly controlled testing conditions. In addition, research to develop evidence-based methods for validating the presence of cell-free fetal DNA would strengthen the test by ensuring that negative test results reflect the presence of a true female fetus and not the absence of fetal DNA. A standard set of biallelic maternal and paternal markers—locations across the genome where defined variation allows for comparison of different genome sequences—could provide this validation. Last, this technology is currently advertised directly to consumers by Internet-based companies, some of which have claimed test accuracies of 95% to 99% at as early as 5 to 7 weeks' gestation.^8- 9 These companies could use the results of such reports to ensure that their claims are accurate. In addition, it would be useful to directly query consumers of these tests to better understand their motivations for undergoing testing as well as their intended or actual actions following receipt of test results.

In summary, the overall performance of noninvasive fetal sex determination using maternal blood can be high, if performed using RTQ-PCR on a blood sample taken at a time during pregnancy when sufficient cell-free fetal DNA is present (7 weeks' gestation or later). This technology can be useful in clinical settings for early detection of fetuses at risk for sex-linked disorders requiring follow-up testing.