Miscarriage is the most common complication of early pregnancy, and it is a particular concern when bleeding is a presenting symptom.1–4 A combination of serial human chorionic gonadotropin (hCG) measurements and ultrasonography has been demonstrated to be the optimal strategy to determine embryonic viability and to ensure that appropriate interventions are performed, especially if there is a concern for ectopic pregnancy.5 The accurate diagnosis of miscarriage is complicated in symptomatic women who present with an ultrasonogram that is initially nondiagnostic.6,7 Serial hCG measurements that are less than the minimal increase (less than 53% over 48 hours)8 may be predictive of a nonviable pregnancy but provide limited discrimination between a miscarriage9 and an ectopic pregnancy.10 Our previous research defined the natural decline in hCG in an urban-dwelling population of women who presented for evaluation of pain and bleeding with an unconfirmed diagnosis at presentation and an ultimate diagnosis of miscarriage.11 In this population, a quadratic curve for log hCG was determined to best fit the pattern of hCG decline. The minimal rate of decline in patients with miscarriage was found to be more rapid for women who presented with higher initial hCG concentrations and ranged from 21% to 35% 2 days after presentation or 60% to 84% at 7 days of follow-up.12 A woman with a decline in hCG slower than this threshold was considered at risk for an ectopic pregnancy.
The primary aim of this research was to redefine the natural decline of hCG in women presenting with a pregnancy of unknown location and with ultimately diagnoses of resolved pregnancy of unknown location. An updated nonlinear mixed-effects model was used to derive curves representing hCG elimination in a geographically and racially diverse national cohort. In addition, demographic variables and risk factors for miscarriage were evaluated for their potential effect on the longitudinal profiles of serial hCG measurements in women who experience miscarriage.
MATERIALS AND METHODS
Clinical data used for this study were collected at three sites, University of Pennsylvania, University of Miami, and University of Southern California, as part of the Predictors of Ectopic Pregnancy Study. Institutional Review Board approval was obtained at each site. Data were collected from September 2007 through May 2009. The cohort for this study consisted of women who presented with pain, bleeding, or both and a positive pregnancy test result. Clinical data from each site were entered into a centralized database that was queried for this investigation. The study was powered to be able to estimate a sensitivity of 80% for the prediction of ectopic pregnancy compared with nonectopic (miscarriage and viable intrauterine gestation), with a 95% confidence interval of ±5%. These calculations also assumed that the outcomes would be distributed as 18.6% ectopic pregnancy, 26% viable intrauterine pregnancy, and 55.4% miscarriage.
To be eligible for study, each woman had to have a nondiagnostic transvaginal ultrasonogram at presentation (no evidence of an intrauterine or extrauterine gestational sac), an initial hCG value of less than or equal to 10,000 milli-international units/mL, and follow-up with serial hCG measurements to determine the definitive diagnosis. The diagnosis of resolved pregnancy of unknown location was defined as a spontaneous decline of hCG to less than 5 milli-international units/mL (the lower limit of detection for the hCG assay used) or at least three declining hCG values, with the final hCG value less than 25 milli-international units/mL. Of note, the data also were censored at hCG values of 10 milli-international units/mL or less, at which point values were treated effectively as undetectable. Only women with declining hCG levels without medical or surgical intervention were enrolled in the study. All information up to 35 days from presentation was included in the analysis and used for creation of the final model.
Serum hCG concentrations were quantified using a chemiluminescent technique (Axysm) and reported in milli-international units/mL using the Third International Reference Preparation (coefficient of variation 7.8). Information concerning last menstrual period, race, ethnicity, and maternal age were collected at the time of initial visit.
Analyses were conducted using the natural log transformation of the hCG values to reduce the skewness of the distribution and to reduce the influence of large values. To reproduce our original article,11 a linear random-effects model was used to estimate the hCG slope per participant along with predictive intervals or upper and lower confidence bounds (percentiles) for those slopes. In addition, we used a nonlinear mixed-effects model that included random effect on the time axis to estimate the average shift at each site and allowed an additional shift by each person relative to the average. This approach, known as curve registration,12 allows the curve to be shifted for each respondent by an amount of days from the presentation date to align the rate and generate the best-fitting overall curves. With the use of this second model, there was no need to know the date of the last menstrual period. The model computed a revised days-from-presentation variable (called “tau” in this article to distinguish it from “time”) using a shift in days for each respondent that was composed of a specific fixed effect for the site plus a random effect attributable to each individual. This new day variable “lined-up” the respondents based on their values and the shape of their individual curves. “Tau” was defined as follows: tau=days from presentation+a shift for site a per-participant shift (ie, random effect).
Using this shift, we tested a linear model and a quadratic model for tau. The linear model was as follows: log_hCG=intercept+b1×tau. The quadratic model was as follows: log_hCG=intercept+b1×tau+b2×tau2.
Versions of these models were compared to determine whether the time shift could account for the effects of sites and whether a linear model would be adequate over a short range of time from presentation. In addition, covariates including race, the presence of bleeding, and other demographic and clinical variables were explored to determine whether different predictive models would be needed for patients with these covariates.
To create tables of predicted hCG values, we first performed the models taking out the fixed effects for the sites. The rationale for this was that when we apply that to the prediction of hCG values for a new site, the effect of that site would be unknown. This slightly increased the variance of the prediction, resulting in slightly wider prediction interval ranges (ie, bounds). We then computed a starting value of tau based on the initial hCG value. Higher hCG values shift a person closer to the presentation date, thus moving them into an area of the curve with a steeper or more rapid decline. Second, we computed the expected change in hCG values across time periods of 1, 2, 3, 4, and 7 days. Third, we computed the “percentile ranges” in these values using estimated standard errors for the slope parameters and percentiles of the per-person shift variable tau. Thus, the decline listed for the 95th percentile means that only 5% of the patients with a resolved pregnancy of unknown location would have a decline in hCG this slow or slower.
To determine the best method for combining the data from the three sites, we created a model with the three sites as parameters. In a second model, we added a time shift for each site. When we added the time shift, the significance level for the three sites was reduced to nonsignificance. To determine whether a linear model for these data would be adequate over a restricted range instead of a quadratic model, we compared the fit of the linear, quadratic, and cubic models using the likelihood ratio test. The quadratic model fit the data better and was used to generate the estimates. The scale of the graph is in the “new days” variable, labeled “tau;” thus, the time scale is not in days from presentation at the clinic, but rather shifted days.
Predicted hCG curves derived from this method determined the minimum expected hCG decline in a resolved pregnancy of unknown location. From a clinical perspective, once the observed rate of decline in hCG becomes less than the minimum defined by the curve, the patient is classified as having a suspected ectopic pregnancy. Two such scenarios for different initial hCG concentrations are represented graphically in Figure 1A and B. All analyses were conducted using SAS 9.2, Proc Mixed, and Nlmixed.
The initial sample included 518 women who presented with a pregnancy of unknown location and had diagnoses of spontaneously resolved pregnancy of unknown location. A total of 33 women were excluded who did not have at least two hCG values separated by less than 7 days, and 42 women were excluded because the initial hCG values were more than 10,000 milli-international units/mL. The final sample included 443 eligible women.
Participant characteristics according to clinical site are presented in Table 1. Overall, 42% of women were Caucasian, 48% were African American, and 10% self-reported “other” as race. Thirty-five percent reported Hispanic ethnicity. The following variables differed significantly by study site: racial and ethnic distribution of women by design; age; initial hCG; number of hCG values collected; days of follow-up; presence of pain at presentation; gestational age at presentation; and certainty of last menstrual period. Study sites were similar with respect to obstetric history of women and bleeding at presentation.
Figure 2 depicts the predicted hCG curve for resolving pregnancy of unknown location based on the model using all data points contributed by members of the cohort. Values corresponding to the median hCG concentrations are flanked by the 95% and 5% values.
The effect of specific variables on the decline of hCG in miscarriage is presented in Table 2. Demographic and clinical factors that significantly modified the hCG curves included age older than 35 years (P=.001) and pain at initial presentation (P=.006). These results, depicted graphically in Figure 3A and B, show that decline in log-transformed hCG is less rapid in women who are older than 35 years at presentation compared with those younger than 35 years. Conversely, hCG decline was more rapid in women who described pain at presentation compared with those who did not.
Table 3 presents the average percentage decline in hCG levels as a function of the initial hCG value across time in our new data compared with what were reported previously by our group.11 The rate of hCG decline is directly proportional to the magnitude of hCG at presentation. On average, with an initial hCG of 2,000 milli-international units/mL and follow-up at 2 days, women experience a decline of at least 58%. The comparable 2-day follow-up value in women presenting with hCG of 5,000 milli-international units/mL was 60%. The average decline in hCG in our new population was slower than previously published (at all initial hCG concentrations). However, mean and median values of the decline were similar (data not shown).
For the purpose of using these curves to distinguish spontaneous miscarriage from ectopic pregnancy, our focus was on the extremes of the predicted curves (90th percentile and 95th percentile) for each starting hCG level that describes the minimum predicted hCG decline and provides a threshold for clinical discrimination. Based on our modeling, the minimum rate of hCG decline at 2 days (95th percentile) of follow-up ranged from 35% to 50% (depending on initial hCG), and ranged from 66% to 87% at 7 days. The values from this analysis are juxtaposed with our previous data in Table 4. Compared with our previously published findings regarding the decline of hCG in women with miscarriage, these new data demonstrate a greater expected minimum decline for those women considered to have a spontaneously resolved pregnancy of unknown location (ie, the 90th and 95th percentiles of the projected curve). For example, with a starting hCG of 2,000 milli-international units/mL, the curve from 2004 predicted that a minimum decline of 31% (95th percentile hCG value) at 2 days was consistent with spontaneous miscarriage,11 whereas the updated curve predicts a minimum decline of 46% (95th percentile hCG value) at 2 days in women with a resolved pregnancy of unknown location. The contrast between the curves is less marked at 7 days.
The goal of this investigation was to present hCG elimination curves in women with a spontaneously resolved pregnancy of unknown location derived from an ethnically and geographically diverse population using innovative methods to derive the curves. The use of linear mixed-effects models to generate an hCG slope and nonlinear mixed-effects models to estimate the shift by each person in the sample better-reflects the natural history of the decline in hCG in this population. The updated analyses demonstrated that whereas the mean hCG decline of a resolving pregnancy of unknown location was slower than previously reported, the threshold (90th or 95th percentile of the project curves) to define when the hCG was not declining “fast enough” to be considered resolved was higher. The difference between the two analyses regarding minimum thresholds to define whether hCG was resolving fast enough (at a given time and given an initial hCG) illustrates the variability of hCG resolution or could be an artifact of our modeling. Differences between current and previous analyses may be explained by competing factors influencing the generation of hCG by remaining trophoblast and the clearance of hCG, and were most pronounced with early hCG follow-up (2 days after presentation). Such variability emphasizes the need to exercise caution when interpreting the decline in a woman with a pregnancy of unknown location, especially when values are close to reported minimum thresholds.
In our current analysis, the 95% predictive bound for the minimum hCG decline in women with resolving pregnancy of unknown location ranged from 35% to 50% at 2 days and from 66% to 87% at 7 days for starting hCG values of 250–5,000 milli-international units/mL. In our previously published data, the minimum rate of hCG decline that was consistent with miscarriage ranged from 24% to 35% at 2 days and from 68% to 84% at 7 days.11 In our study validating the concept of comparing hCG curves of women at risk for ectopic pregnancy with that of expected normal women with resolved pregnancy of unknown location or viable intrauterine pregnancy, the threshold that misclassified the fewest ectopic pregnancies was an hCG decline of 36–47% at 2 days, which is closer to our current estimates.13 However, it was noted in that study that a high number of women with miscarriage would have undergone intervention to maximize the number of women with ectopic pregnancy diagnosed.
A decline in hCG slower than the predictions we report herein identifies a woman at risk, but it does not make the diagnosis of ectopic pregnancy. Continued surveillance is warranted with additional values, especially when a woman remains pain-free. The number of instances when the pattern of hCG can mislead a clinician into a false diagnosis can be minimized by obtaining additional values as well as considering the entire clinical situation.13 Even when a pregnancy is considered nonviable, the use of interventions such as uterine curettage and methotrexate administration should be reserved for management when clinical indicators accumulated over a reasonable interval suggest an increasing concern for ectopic pregnancy. Thus, the optimal threshold to aid in the diagnosis of ectopic pregnancy and miscarriage is an hCG decline of approximately 35–50% at 2 days but is dependent on the clinical situation, including acceptance of intervention.
We also report that clinical and demographic factors affect the decline in hCG (Fig. 3). One possible explanation for our findings of more rapid hCG decline in women with a pregnancy of unknown location experiencing pain is that increased clearance of hCG occurs because of greater uterine contractility, resulting in extrusion of uterine contents. It is not readily apparent why maternal age should affect hCG elimination curves in a pregnancy of unknown location. The primary determinant of hCG half-life is the carbohydrate content (sialic acid) of the hormone that is directly proportional to its biological activity.14 Although it is possible that older women have differential patterns of hCG glycosylation that result in diminished metabolism of the hormone, data supporting this could not be found in the literature. However, because the effects of age and pain on hCG resolution are subtle, and because of the large overlap in hCG values, these variables appear to have limited clinical utility. The most important effect modifier of the hCG decline remains initial hCG value.
The current hCG curves are derived from a population of women that is more diverse, geographically, ethnically, and racially, than in our previous report, which is a major strength of this research. Another strength of our findings is the completeness of data collection that allows for the availability of expected declining hCG levels at multiple time points up to 7 days after presentation.
In summary, we have presented newly derived curves that predict the decline in hCG in women presenting with a pregnancy of unknown location and ultimate diagnoses of resolved pregnancy of unknown location. Declines less steep at characterized time points warrant further evaluation to differentiate miscarriage from ectopic pregnancy. These updated data, derived from the best methods available, should enhance the discriminatory power of the hCG curves to guide clinical care. Ultimately, the rules that define declining hCG in miscarriage must be incorporated into an approach to care that includes best clinical judgment, an appropriate number of serial hCG assessments, and risk factor assessment to provide an accurate and timely final diagnosis for women presenting with a pregnancy of unknown location.
1. Deutchman M, Tubay AT, Turok D. First trimester bleeding. Am Fam Physician 2010;81:577.
2. Snell BJ. Assessment and management of bleeding in the first trimester pregnancy. J Midwifery Women's Health 2009;546:483–91.
3. Zinaman MJ, Clegg DE, Brown CC, O’Connor J, Selevan SG. Estimates of human fertility and pregnancy loss. Fertil Steril 1996;65:503–9.
4. Ellish NJ, Saboda K, O’Connor JO, Nasca PC, Stanek EJ, Boyle C. A prospective study of early pregnancy loss. Hum Reprod 1996;11:406–12.
5. Gracia CR, Barnhart KT. Diagnosing ectopic pregnancy: decision analysis comparing six strategies. Obstet Gynecol 2001;97:464–70.
6. Barnhart KT, Fay C, Suescum M, Sammel MD, Appleby DH, Shaunik A, et al.. Clinical factors affecting the accuracy of ultrasound in symptomatic first trimester pregnancy. Obstet Gynecol 2011;117(2 Pt 1):299–306.
7. Barnhart KT, Simhan H, Kamelle SA. Diagnostic accuracy of ultrasound above and below the beta-hCG discriminatory zone. Obstet Gynecol 1999;94:583–7.
8. Barnhart KT, Sammel MD, Rinaudo P, Zhou L, Hummel AC, Guo W. Symptomatic patients with an early viable intrauterine pregnancy: hCG curves redefined. Obstet Gynecol 2004;104:50–5.
9. Barnhart K, van Mello NM, Bourne T, Kirk E, Van Calster B, Bottomley C, et al.. Pregnancy of unknown location: a consensus statement of nomenclature, definitions and outcome. Fertil Steril 2011;95:857–66.
10. Chung K, Sammel MD, Zhou L, Hummel A, Guo W, Barnhart K. Defining the curve when initial levels of human chorionic gonadotropin in patients with spontaneous abortions are low. Fertil Steril 2006;85:508–10.
11. Barnhart KT, Sammel MD, Chung K, Zhou L, Hummel AC, Guo W. Decline of serum hCG and spontaneous complete abortion: defining the normal curve. Obstet Gynecol 2004;104:975–81.
12. Ramsey JO, Li X. Curve registration. J Royal Stat Assoc Series B 1998;60:351–63.
13. Morse CB, Sammel MD, Shaunik A, Allen-Taylor L, Oberfoell N, Takas P, et al.. Performance of human chorionic gonadotropin curves in women at risk for ectopic pregnancy: exceptions to the rules. Fertil Steril 2012;97:101–6.
© 2013 by The American College of Obstetricians and Gynecologists.
14. Frtiz M, Speroff L. Clinical gynecologic endocrinology and infertility. 8th ed. Philadelphia (PA): Lippincott Williams and Wilkins; 2010.