An accurate method to predict subsequent miscarriage in live embryos has not yet been established. This pilot study aimed to determine the most discriminatory ultrasound-based model for predicting spontaneous miscarriage after embryonic life was first detected in assisted conceptions. A method for estimating individual risk of miscarriage was developed.
This was a prospective cross-sectional survey of 322 live singleton embryos in women from an assisted reproductive technology program. Mean sac diameter (MSD), crown–rump length (CRL), embryonic heart rate (EHR), maternal age and gestational age at the first transvaginal scan detecting embryonic life (between 42 and 62 days) were observed. These variables were included in a multivariate model for predicting spontaneous miscarriage occurring prior to 20 weeks. MSD, CRL and MSD minus CRL were assessed in univariate logistic regression analyses. The global diagnostic accuracy of each model was compared directly using receiver–operating characteristics (ROC) curves.
The multivariate model demonstrated the best ROC curve for predicting miscarriage (ROC area 0.87; 95% CI, 0.80–0.95). The separate univariate analyses had less diagnostic accuracy. In particular, MSD − CRL had a significantly smaller ROC area (0.65) than did the multivariate model (P < 0.01).
An accurate model for the prediction of subsequent miscarriage occurring in early live pregnancies has not been established. The first detection of embryonic heart motion by ultrasound does not ensure ongoing pregnancy viability, even in the absence of symptoms. The incidence of spontaneous miscarriage occurring after embryonic heart motion has been detected sonographically is high, ranging from 5 to 20%1–4. The ability to make an accurate risk estimate for the prediction of subsequent pregnancy failure is therefore important not only for counseling individual prospective parents, but also from the broader public health perspective. Multiple ultrasound examinations are costly and an accurate estimate of the likelihood of spontaneous miscarriage would serve to rationalize the use of additional scans.
A high sensitivity for predicting subsequent miscarriage in early live embryos has been demonstrated when the difference between the size of the gestational sac and embryonic pole is reduced. This ultrasound criterion was first described in six early live embryos that subsequently all miscarried, each with a normal crown–rump length (CRL) but all with small gestational sac volumes that were two SDs or more below normal5. A more current and widely accepted measure of gestational sac size, the mean sac diameter (MSD) has been used to reproduce this work6; a small MSD predicted first-trimester miscarriage in 94% of cases when the difference between the MSD and CRL (MSD − CRL) was less than an arbitrary threshold of 5 mm. These early series were small and did not take precise gestational age data into account. Furthermore, there is controversy regarding the use of MSD − CRL as the primary criterion in predicting early pregnancy prognosis. A prospective study of 39 late first-trimester miscarriages concluded that in spite of using a larger MSD − CRL discrepancy threshold of 10 mm, this criterion was far less sensitive for predicting miscarriage than when both the MSD and CRL measurements were smaller than expected for the gestational age: 56% vs. 71% sensitivity, respectively7. None of these series considered embryonic heart rate (EHR) as an additional factor in predicting pregnancy outcome.
Irrespective of its performance, the MSD − CRL model has established popularity as a simple and objective method for predicting miscarriage most likely because the calculation requires no reference to tables or nomograms for each individual pregnancy. As a previously established means of early pregnancy prognostication the MSD − CRL model is also well suited to comparative analyses primarily because of its objective and quantitative nature which allows for direct comparisons with future models. This is in contrast to subjective methods of ascribing pregnancy prognosis such as irregular gestational sac shape and low sac position.
Our study was designed to critically evaluate the significance of the MSD − CRL model compared with a new multivariate model for the prediction of subsequent miscarriage in a population of pregnant women with live singleton embryos conceived using assisted reproductive technology (ART). The proposed multivariate model combined basic ultrasound parameters with known risk factors for miscarriage such as maternal age and EHR. This prospective study of 322 live singleton embryos, all with accurate gestational age estimates from ART data, was performed as a pilot study to determine the best model for establishing early pregnancy prognosis. The use of logistic regression analyses of individual univariate models, including MSD − CRL, allowed a direct comparison of this previously established model with the present multivariate model in order to identify the most discriminatory test for predicting subsequent miscarriage in early live embryos. This is the first study to make such a direct comparison with precise gestational age data available, in a population not selected for bleeding symptoms. In addition, a new method for generating individual risk figures for subsequent miscarriage based on the present multivariate model is proposed.
This was an observational study of 322 women with live singleton pregnancies conceived on an ART program from January 2000 to April 2001. The study design was cross-sectional and only data from the single scan at which embryonic heart motion was first detected for each of the pregnancies was included. The scan data from each pregnancy were entered consecutively without selection for symptoms of threatened miscarriage.
The ART pregnancy protocol was as follows. By convention estimated menstrual age was used and calculated from the date of embryo transfer, taking into account the age of the embryos used. Progesterone support was not part of standard protocol. The women were assessed for quantitative serum β-hCG and progesterone levels on days 16 and 23 post embryo transfer.
Ultrasound scans were indicated when there was biochemical evidence of an ongoing pregnancy. The first scans were performed transvaginally in order to identify embryonic life at various gestational ages from as early as 42 days onwards as directed by their specialist. Most patients were scanned at 6–7 weeks. If no live embryo was detected at the initial scan then a scan was repeated after a 7-day interval. A second scan was required to identify a live embryo in 0.93% of pregnancies (3/322). At the time of the first scan, the following background data were obtained: maternal age, prior history of miscarriage and symptoms of bleeding in early pregnancy. The precise gestational age was recorded from the available ART treatment cycle data and expressed as estimated menstrual age by convention. The ultrasound parameters recorded at the time of this first scan establishing embryonic life were MSD, CRL and EHR. The MSD was calculated in mm from the average of three orthogonal dimensions measured from the inner sac wall/chorionic fluid interface. The CRL was taken as the greatest length in mm of the embryo. The difference between the MSD and CRL in mm was then calculated for each case.
Pregnancies were excluded from the study if there was evidence of more than one gestation or if no live embryo was identified. No pregnancies in women with uterine anomalies were included in this sample. A technical team of six dedicated obstetric sonographers supervised by two full time sub-specialists in obstetric and gynecological ultrasound performed all the transvaginal ultrasound examinations. The scans were carried out on ATL HDI 3000 and 3500 systems (Phillips Ultrasound, Bothell, WA, USA), using 8–4-MHz transvaginal transducers. All measurements were taken according to established unit protocols. Interobserver variation was minimized by standardization of the individual techniques against the primary researcher prior to commencement of the project. Thermographic prints of each study were submitted to the supervising sub-specialists to enable ongoing technical audit.
The pregnancies were followed for survival to 11–14 weeks and to 18–20 weeks by review of subsequent ultrasound scan results. If no subsequent scan results were available, then clinical pregnancy outcome data was sought from the treating obstetrician. The adverse outcome variable was spontaneous miscarriage, which was defined as pregnancy loss occurring subsequent to the first viability scan and prior to 20 weeks. This complies with the clinical definition for spontaneous miscarriage in Victoria, Australia. The gestational age at the time of the pregnancy loss was recorded.
Data were collected on Microsoft Excel 2000 (Microsoft Corp, Redmond, WA, USA) and statistical analyses were performed using the Stata statistical package version 5 (Stata Corporation, College Station, TX, USA). Individual univariate analyses using logistic regression were performed separately for MSD and CRL and MSD − CRL to examine the association with spontaneous miscarriage prior to 20 weeks as a binary outcome. The estimates for parameters for each model were expressed as odds ratios (ORs) and their respective 95% confidence intervals (CIs).
A multivariate model was then developed using logistic regression to examine the association with spontaneous miscarriage, including five variables in combination: maternal age, gestational age at the scan when embryonic heart motion was first detected, MSD, CRL and EHR. This produced the ORs of spontaneous miscarriage based on each observed variable, correcting for the confounding effect of the other variables included in the multivariate analysis. In particular, including gestational age as a factor in the multivariate analysis accounts for any confounding effect of this variable on EHR. MSD − CRL was not included as a parameter in the multivariate model because this value is a function of MSD and CRL and therefore strongly correlated with both these variables. The results of the multivariate analysis, and the univariate analyses for MSD only and CRL only were compared with those of the univariate analysis for MSD − CRL.
A separate receiver–operating characteristics (ROC) curve was generated for four models: the MSD only, CRL only, MSD − CRL and the multivariate models. The ROC curves of the MSD only, CRL only and multivariate models were compared with that of the MSD − CRL model using the area under the curve8 (ROC area); this area was used as a global assessment of the performance or diagnostic accuracy of each model for spontaneous miscarriage. P-values for determining if significant differences existed between the ROC areas of each of these models versus that of the MSD − CRL model were also estimated. The regression coefficients and 95% CIs for each parameter were calculated and a method for estimating the individual risk of miscarriage based on the regression formula for the multivariate model was produced.
Observations were made on a total of 322 live singleton pregnancies during the 16-month course of this study. Outcome data on pregnancy survival or spontaneous miscarriage were available for 301 (93%) of these pregnancies, which formed the basis of our analyses. The age distribution of the study population was typical for an ART population, with a mean (SD) maternal age of 34 (4.4) years. The gestational age at which the viability scans were performed ranged from 42–62 (median, 44; SD, 3.8) days. The overall incidence of early pregnancy bleeding in the study population was 19.3% (58/301) and 7.9% (24/301) had a past history of spontaneous miscarriage. Spontaneous miscarriage prior to 20 weeks occurred in 13.6% (41/301) of cases, at a mean gestational age of 10 (range, 6–15) weeks. The incidences of both early pregnancy bleeding and past history of spontaneous miscarriage were no higher in the miscarriage group compared with the overall study population. Of the 41 women who suffered spontaneous miscarriage, seven (17%) had bleeding symptoms and only two (4.8%) had a past history of miscarriage.
The results of the separate logistic regression analyses are presented in Table 1. Univariate analysis of the MSD − CRL model demonstrated a significant effect for the prediction of miscarriage (OR, 0.85; 95% CI, 0.77–0.94). MSD and CRL as separate parameters demonstrated similar significant associations with subsequent miscarriage in univariate analyses as well as when part of a multivariate model. The ROC curves for each model are depicted in comparison with that of the MSD − CRL model in Figures 1–3. Table 2 sets out the comparative results of the MSD only, CRL only and multivariate model against the previously established MSD − CRL model. These comparisons showed that there was a statistically significant difference in the ROC areas between the MSD only and the MSD − CRL models (P < 0.01). MSD as an individual parameter was a better overall predictor of spontaneous miscarriage compared with MSD − CRL (ROC area, 0.75 vs. 0.65). There was an even greater difference in the ROC areas between the multivariate model and the MSD − CRL model which was also statistically significant (ROC area, 0.87 vs. 0.65; P < 0.01). The best diagnostic accuracy for spontaneous miscarriage was therefore achieved when five variables unique to each woman's first viability scan were included in a multivariate model (i.e. maternal age, gestational age at viability scan, MSD, CRL and EHR). The regression coefficients in the multivariate predictive model for spontaneous miscarriage are shown in Table 3. From this the following equation was developed to create an individualized risk assessment for singleton pregnancies following assisted reproduction.
Table 1. Comparison of odds ratios (ORs) of models predictive for spontaneous miscarriage
Quadratic terms for all variables were tested and found to be non-significant.
Gestational age at which live embryo first detected. CRL, crown–rump length; EHR, embryonic heart rate; MSD, mean sac diameter.
Equation for estimating the individual risk of subsequent miscarriage in early live embryos using regression coefficients for the multivariate model from Table 3
To use the parameters in the multivariate model, X* must be calculated by the formula:
X* is then transformed into the predicted probability via the equation:
For example, consider the individual who has age = 43 years, gestational age at viability scan = 45 days, MSD = 10 mm, CRL = 7 mm and EHR = 100 bpm. X* is:
and the probability of miscarriage is:
The aim of this study was to identify the model with the best diagnostic accuracy for predicting subsequent spontaneous miscarriage in women with live singleton embryos conceived using ART. The previously established MSD − CRL model was compared with univariate models of MSD, CRL and a new multivariate model. The multivariate model was based on maternal age, gestational age, MSD, CRL, and EHR data obtained at the first transvaginal scan that established the presence of a live embryo. Analyses of the data presented from this ART population clearly demonstrated that the multivariate model had the best ability to discriminate between pregnancies destined to abort spontaneously and those that survived into the second trimester. Specifically, this multivariate model had a significantly better ROC curve compared with that obtained using MSD − CRL as the sole parameter for determining pregnancy prognosis (Table 2 and Figure 3). MSD as an individual parameter also performed better in a comparative univariate analysis than did the MSD − CRL model in predicting miscarriage (Table 2). Our data did not therefore show any diagnostic advantage to support the use of MSD − CRL as a predictive parameter.
We included MSD and CRL as parameters in the current multivariate predictive model as both of these variables are expected to increase with advancing gestational age consistent with early growth in both the chorionic sac and the embryonic pole. Separate univariate analyses of these parameters demonstrated a reduction in the risk of miscarriage as each of these variables increased (Table 1). The same effect was shown for each of these parameters within the multivariate analysis. Unique non-ultrasound-based factors also have a role in predicting miscarriage. EHR and maternal age were included as factors within our multivariate predictive model because these are recognized risk factors for miscarriage. Embryonic bradycardia of 85 bpm or less is universally associated with impending miscarriage9–11. A trend of decreased miscarriage risk was demonstrated in association with increasing EHR within our multivariate analysis (OR, 0.96; 95% CI, 0.92–1.00). It has long been recognized that advanced maternal age is a risk factor for early pregnancy loss. The biological explanation includes the fact that meiotic non-disjunction during oogenesis is more frequent in older mothers and their offspring are at increased risk of trisomies 13, 16, 18 and 2112, 13. The incidence of chromosomal abnormalities in spontaneous abortuses during the first trimester has been reported to be as high as 61.5%14. In the current multivariate analysis, increasing maternal age demonstrated a trend towards an increased risk of subsequent miscarriage as expected (OR, 1.14; 95% CI, 1.03–1.27).
In the setting of the present study, the precise gestational age at which a live embryo was first detected for each pregnancy was available from the individual ART treatment cycle data. Transvaginal ultrasound scans were routinely performed in these pregnancies from 42 to 62 days' gestation in order to identify a live embryonic pole. In the multivariate analysis, the later in gestation embryonic heart motion was first seen, the higher the risk of subsequent miscarriage (OR, 1.75; 95% CI, 1.41–2.17). This is presumably because the less healthy embryos destined to eventually abort are smaller and therefore more difficult to visualize on transvaginal ultrasound until later in gestation.
The theoretical advantage of a predictive model based simply on MSD − CRL is that knowledge of accurate gestational age data is not required. The evidence supporting the use of an abnormally small difference in size between the gestational sac and the embryonic pole as a test for subsequent miscarriage comes mainly from the earliest series5, 6. An arbitrary MSD − CRL threshold of 5 mm or less has been used to define abnormally small sacs in a prospective series6. By this criterion the estimated risk of subsequent miscarriage was as high as 94% (15/16) vs. 8% in 52 controls, if the sacs were small in the presence of a normal EHR. This study was, however, small and it did not attempt to assess the size of either the sac or the embryo as individual parameters with respect to gestational age.
The performance of MSD − CRL as the primary criterion for establishing early pregnancy prognosis is still controversial. A larger prospective series using an even lower MSD − CRL threshold of 10 mm demonstrated a much lower sensitivity of only 56% for predicting miscarriage7. The authors compared the predictive ability of small-for-dates MSD and CRL (with respect to known conceptual dates), vs. that of MSD − CRL and found a higher sensitivity for predicting miscarriage using either small sac size (82%) or small embryo size (77%) as individual parameters. This is consistent with our data. The sensitivity for miscarriage when both the MSD and CRL were smaller than expected for known gestational age was 71%, again higher than that for MSD − CRL. These findings support the use of MSD, CRL and accurate gestational age data as separate parameters within a multivariate model for the prediction of subsequent miscarriage in early live embryos.
Data from more recent series further support the practice of including multiple parameters for predicting subsequent miscarriage in sonographically established live embryos. The first study to combine gestational sac size (mean of two diameters), CRL and abnormal EHR for the prediction of early pregnancy outcome was a large series of 603 embryos15. Twenty-three pregnancies ended in first-trimester miscarriage after viability had first been established sonographically. The authors did not use the traditional MSD − CRL parameter; instead, nomograms were generated for the ratio of sac size/CRL against EHR, gestational age and CRL. They found that combining abnormal sac size/CRL ratios with abnormal EHR values based on arbitrary thresholds had a higher sensitivity for miscarriage (78.3%) than did either parameter alone. Their two-dimensional sac size/CRL ratio is not however an established parameter of embryonic growth and it was unfortunate that the authors did not use the conventional three-dimensional MSD which is a more widely accepted parameter of sac size. Direct comparisons of results from other series are therefore extremely difficult to make.
Logistic regression was used to study prediction of spontaneous miscarriage in 149 viable early pregnancies with bleeding16. In this prospective study a significant association was found between the occurrence of miscarriage and fetal bradycardia, an abnormal MSD − CRL (< 0.5 SD from the mean) and discordant menstrual and sonographic age (> 1 week). The probability of miscarriage was highest when all three risk factors were present (84%). In this comparative study, embryonic bradycardia had the most powerful independent association with pregnancy outcome in women with threatened miscarriage. No direct comparison was made, however, between a predefined multivariate model and one based solely on MSD − CRL for predicting spontaneous miscarriage. Our study had a larger sample size and the benefit of precise gestational age estimates. Also, the sample population in our study was not selected on the basis of bleeding symptoms in order to better reflect the wider ART population.
We acknowledge that the predictive implications of the present pilot study are confined to an ART population of older women with precise gestational age data available, and may be applicable only in the gestational age range of 42–62 days. Our multivariate equation requires prospective evaluation in a new sample population. Further study would be required to establish the optimum risk threshold for detecting miscarriage, which would include consultation with the caring physicians. Accepting that no screening test is completely accurate, then the optimum risk threshold of a predictive model is when the test-positive cut-off risk yields a specific combination of sensitivity and specificity that is acceptable clinically. In this setting, for example, a false-positive result would result only in exposure to a repeat scan so a cut-off risk with a relatively high false-positive rate could conceivably be acceptable in order to produce a higher sensitivity. Including a wider gestational age range may make this model more relevant. The predictive equation may also require adjustment for populations with different maternal or gestational age ranges depending on the referral base and ultrasound protocols of specific departments.
There is a clear clinical advantage in using a multivariate model for predicting early pregnancy outcome. Based on the multivariate regression equation and respective coefficients from analysis of our data (Table 3), the probability of subsequent spontaneous miscarriage can be generated for each individual singleton ART pregnancy from parameters assessed at the first scan demonstrating a live embryo. By evaluating the individual risk of spontaneous miscarriage against the background population risk, any follow-up scans can be directed more effectively. There is thus a potential not just for improving clinical practice but also for decreasing public health expenditure. On an individual note, any reassurance or caution regarding pregnancy outcome may also be given based on a more robust predictive model. This is especially helpful for couples with assisted conception as there is often a high background level of anxiety.
Further study in a new ART population is now required to test the performance of this model as a screening test for subsequent miscarriage, assessing for sensitivity, specificity and predictive values. The optimum risk threshold for predicting pregnancy outcome also needs to be established.
We wish to thank the staff of Monash Ultrasound for Women. Leanne Elliot, Karen Waalwyck, Justine Armstrong, Helene Johns, Annette Fry and Andrea Materia are acknowledged for their contribution to data collection.