Diagnostic value of serum hCG on the outcome of pregnancy of unknown location: a systematic review and meta-analysis

Abstract

BACKGROUND

The term ‘pregnancy of unknown location’ (PUL) refers to cases where a pregnancy test is positive but the pregnancy cannot be visualized by transvaginal sonography (TVS). Various strategies integrating TVS and serum hCG measures are used to follow-up until the location and/or viability of the pregnancy becomes clear; however, the optimal strategy to predict the outcome of pregnancy in women with PUL is unknown. Therefore, we performed a systematic review and meta-analysis to determine the diagnostic accuracy of the various serum hCG strategies in women with PUL.

METHODS

We searched Medline and EMBASE for articles which were published (in any language) from 1980 to January 2012 on strategies using serum hCG in women with PUL and reporting on the final outcome of pregnancy.

RESULTS

From 980 selected titles, 23 articles, all cohort studies, were included. There were 10 studies on a single serum hCG cut-off level, 4 on serum hCG ratio (hCG 48 h/hCG 0 h) and 6 on logistic regression modelling. Three other strategies were reported using serum hCG, serum progesterone and/or uterine curettage findings; each of these strategies comprised a single study. Comparative diagnostic studies have not been performed on the diagnostic value of serum hCG in women with PUL. Included studies showed substantial clinical heterogeneity in the definition of the outcome, and only data for the pregnancy outcome ectopic pregnancy (EP) were suitable for meta-analysis. The receiver operating characteristic curves showed that the serum hCG ratios and logistic regression models had a better performance as compared with an absolute single serum hCG level (as the curve was considerably closer to the diagonal, indicating no diagnostic value).

CONCLUSIONS

Overall the study was limited by the high clinical heterogeneity of the data but in women with PUL diagnostic strategies using serum hCG ratios, either alone or in logistic regression models, have the best diagnostic performance in the case of EP. Well defined prospective comparative studies using standardized diagnostics and clinical application plus agreed definitions of outcome are required to identify the best strategy to diagnose pregnancy outcome in women with PUL.

Background

‘Pregnancy of unknown location’ (PUL) is the term used to define a pregnancy which cannot be visualized by transvaginal sonography (TVS). PUL occurs in 7–30% of women who present with complaints in early pregnancy, e.g. abdominal discomfort and/or vaginal bleeding (Kirk et al., 2009).

The majority of women with PUL (50–70%) have a spontaneously resolving pregnancy with serum hCG levels declining to undetectable levels, the so-called failed PUL. Such a pregnancy can either be a failed intrauterine pregnancy (IUP) or a resolved ectopic pregnancy (EP), as the location of the pregnancy remains undetermined. In some women the pregnancy duration is simply too short to allow its visualization on the initial scan. Follow-up scans in combination with rising serum hCG levels will eventually demonstrate an IUP. In 7–20% of women with a PUL, an EP is eventually diagnosed and these women are eligible for laparoscopic surgery or medical therapy with systemic methotrexate (MTX). Only a minority of women will have a persisting PUL, defined as an inconclusive TVS in combination with a rise or plateau in serial serum hCG levels. The optimal management for persisting PUL is not known. Systemic MTX as well as expectant management are reported to be successful (Condous et al., 2003).

There is worldwide consensus that TVS and quantitative serum hCG measurement are useful and should be applied to diagnose the pregnancy location, i.e. the pregnancy outcome in women with PUL (Dutch Society of Obstetrics and Gynaecology, 2001; American College of Obstetricians and Gynaecologists, 2008; Barnhart, 2009; Royal College of Obstetricians and Gynaecologists, 2010). A variety of algorithms have been proposed, integrating clinical examination, TVS and (repeated) serum hCG measurements, to predict the final location and viability of the pregnancy (Cacciatore et al., 1990; Stovall et al., 1992; Ankum et al., 1995; Hajenius et al., 1995; Gracia and Barnhart, 2001; Mertz and Yalcinkaya, 2001; Kohn et al., 2003; Condous et al., 2004, 2005a; Timmerman, 2004). However, no evidence exists on the optimal strategy to predict the outcome of pregnancy in women with PUL. We therefore performed a systematic review and meta-analysis to determine the diagnostic accuracy of the various serum hCG strategies in women with PUL.

Methods

Identification of studies

We performed a literature search using a structured predefined search string on Ovid MEDLINE and EMBASE from 1980 to January 2012. An information specialist performed the electronic search using the terms: ‘unknown/unidentified/undetermined/uncertain pregnancy’ or ‘ectopic pregnancy’, ‘chorionic gonadotrophin’, ‘diagnosis/prognosis/prediction’. We checked cross-references of eligible papers to identify papers not captured by electronic searches. There was a restriction concerning publication year because the introduction of TVS was around the mid-1980s. No restrictions were made based on language.

Selection of studies and data extraction

Two reviewers (N.M. and F.M.) evaluated potentially eligible papers in a two-stage process. First, papers identified in the search were screened for eligibility by reading the title and abstract. If there were any doubts about eligibility, the full text was screened to make sure that no papers were missed. We then obtained full-text versions of all papers which were selected by at least one of the reviewers in the first stage.

Papers were included if the studies reported on women with PUL, serum hCG levels and the final outcome of the pregnancy (IUP, EP, treated persistent PUL or failed PUL). All studies were scored on methodological quality according to the standard for reporting of diagnostic accuracy guideline and the quality assessment tool of diagnostic accuracy studies (QUADAS) and entered on a data extraction form (Khan et al., 2004; Whiting et al., 2006). We scored each study on the following characteristics: study design, data collection (prospective or retrospective), sampling (consecutive women or other) and reported TVS findings. The index test was serum hCG. The reference test differed per pregnancy outcome. To evaluate potential verification bias, we assessed the predefined reference standard as a composite of IUP, EP or other outcomes, as defined by the authors for the QUADAS scoring. Item number 4 of the QUADAS tool, concerning disease progression bias, was removed from the scoring list because the actual pregnancy outcome can only be determined when the pregnancy progresses in women with PUL. We also removed item number 10 from the QUADAS tool because time is an indispensable factor in the diagnosis of women with a PUL and the reference standard was, by definition, performed after the index test.

Studies were included when data were available, either directly or retractable, in which absolute serum hCG levels were expressed for the final pregnancy outcome and/or two-by-two tables for specified strategies were provided.

Statistical analysis

We calculated sensitivity and specificity with 95% confidence intervals (CIs) for each study individually, created forest plots to explore heterogeneity for sensitivity and specificity and plotted their combined results in receiver operating characteristic (ROC) space (sensitivity versus 1 − specificity). To obtain summary estimates of sensitivity and specificity and their 95% CIs and to construct summary ROC (sROC) curves bivariate regression analysis was used (Reitsma et al., 2005; Hamza et al., 2008). The bivariate regression model statistically incorporates the negative correlation that might exist between sensitivity and specificity, because sensitivity and specificity within a study are often negatively correlated, owing to implicit variation of threshold values (Leeflang et al., 2008). To determine the variation or statistical heterogeneity between the results of the studies the random-effects approach was applied. The random-effects approach estimates and incorporates the amount of between-study variability in both sensitivity and specificity. The included studies often reported a different threshold value to define a positive pregnancy test result. To evaluate accuracy measures over the whole range of reported hCG thresholds in all studies, we did not limit our analysis to a single threshold value but estimated accuracy measures for all reported threshold values by assuming that the shift in accuracy (higher sensitivity and lower specificity) due to different thresholds is accounted for by the correlation term, as specified in the bivariate model.

The recommended bivariate modelling approach cannot appropriately account for covariates (such as threshold) with multiple observations from the same study (Leeflang et al., 2008). In order to avoid results being biased towards studies reporting on many different thresholds, we used a stratified bootstrap approach to statistically incorporate different thresholds from the same study in the model. We estimated each model in 50 stratified bootstrap samples, in which only one accuracy estimate from each study was randomly selected. The average overall estimates from 50 bootstrap samples is used to estimate sROC curves, where the increase in sensitivity and decrease in specificity reflect the shift in the threshold value of serum hCG in the model. As the data points represent different thresholds defining a positive test result, the sROC points for these models only reflect the average operating point but have no evident clinical meaning. For the analysis the statistical software package SAS 9.3 (SAS Institute, Cary, NC) was used. A value of P < 0.05 was considered statistically significant.

Results

Number of retrieved papers

Our search retrieved 980 citations from MEDLINE and EMBASE. The process of selection of papers is summarized in Fig. 1. We retrieved no papers from cross-references. After screening titles, abstracts and cross-references, 180 papers were selected for full text review. Of these, 23 studies were included for the meta-analysis.

Characteristics of included studies

The characteristics of the 23 included studies are presented in Table I. Different types of diagnostic strategy using serum hCG were identified. Studies from 1995 to 2011 used a single absolute serum hCG cut-off level (n = 10), studies from 1995 to 2008 used serial serum hCG levels over 48 h, thereby focusing on the hCG ratio (hCG 48 h/hCG 0 h) (n = 4), while more recently (2006–2010) logistic regression modelling strategies were described (n = 6). These six logistic regression models all used serum hCG ratios but combined with different variables: clinical symptoms (n = 2), TVS findings (n = 4), maternal age (n = 2), gestational age (n = 2), risk factors for EP (n = 1) and progesterone levels (n = 4). One study used the prediction by the nurse practitioner combined with hCG ratios, and another the logarithm of the hCG average, the hCG ratio and its quadratic effect. Three other strategies were reported using serum hCG, serum progesterone and/or uterine curettage findings; each of these strategies comprised a single study.

In all studies the definition of PUL was at least a positive pregnancy test and an inconclusive TVS in hemodynamically stable women. Five studies strictly specified the inclusion criteria to women with vaginal blood loss and/or abdominal pain. The other studies included women with PUL, either with vaginal blood loss and/or abdominal pain and/or risk factors for EP and women who came in for first trimester screening without symptoms or risk factors but with an uncertain gestational age. Further specific inclusion criteria among studies were serum hCG cut-off levels varying between 1500 and 10 000 IU/l and an abnormal rise/fall/plateau of serum hCG with an with initial hCG <2000 IU/l.

Of the 23 included studies, 15 were prospective cohort studies and eight were retrospective cohort studies. Not all studies reported on all pregnancy outcomes as the reference test. IUP was variously described as an intrauterine gestational sac, an abnormal sac on TVS, histological confirmation of villi after uterine curettage, spontaneously declining serum hCG without intervention (biochemical pregnancies), fetal heartbeat upon TVS or delivery of a child. Only four studies (from the same author) reported both viable and non-viable IUP separately in the same cohort (Dart et al., 1997, 1999a, b; Dart and Howard, 1998).

For the outcome EP, a composite reference test was used, i.e. confirmation by laparoscopy and additional histology or an ultrasound diagnosis or specific serum hCG cut-off levels. Ultrasound diagnosis ranged from some free fluid in the pouch of Douglas, an inhomogeneous mass adjacent to the ovary, and moving separately from the ovary (blob sign), a mass with a hyper-echoic ring around the gestational sac (bagel sign) (Condous et al., 2005c) to an ectopic gestational sac with or without fetal cardiac activity. Other criteria used to define the outcome as EP were rising or plateauing serum hCG levels after uterine curettage, negative histopathology after uterine curettage, or plateauing serum hCG levels for which systemic MTX was used.

Some studies reported other pregnancy outcomes, such as trophoblast in regression (TIR), failing PUL or miscarriage.

The quality assessment tool (QUADAS) for the included studies is shown in Table II, and from this table it becomes clear that all studies were prone to potential bias. All studies used diagnostic strategies that include the outcome of the serum hCG to determine the type of reference test, e.g. laparoscopy in case of certain serum hCG cut-offs. Thus, the reference standard was not the same for all patients regardless of the index test (differential verification bias) and the reference standard was not independent of the index test (incorporation bias). For the same reason the results of the reference standard were not interpreted without knowledge of the results of the index test (blinding or review bias).

Diagnostic performance of serum hCG in PUL

Sensitivities and specificities, as calculated from each study reporting on serum hCG, are reported for each pregnancy outcome in Table III. We intended to perform a meta-analysis if the same diagnostic serum hCG strategy and the same outcome definition was reported per pregnancy outcome. For the outcome IUP there was sufficient clinical heterogeneity between studies (being either any form of IUP at TVS or specific viable or non-viable IUP) that forest plots or sROC plots would not be informative. Other pregnancy outcomes, such as TIR or failing PUL, were only reported in a few studies and, when reported, the studies did not perform the same diagnostic test. Only the pregnancy outcome EP was reported more consistently, in terms of outcome definition, the reference test and reported data (2 × 2 tables) to allow meta-analysis.

Nonetheless, the accuracy reported for different strategies showed considerable heterogeneity owing to the use of different hCG thresholds (cut-off) to define a positive pregnancy test. We therefore compared the sROC curves rather than the pooled estimates for sensitivity and specificity, as this summary point estimate does not reflect the accuracy for a particular operating point of the test. The pooled data for the pregnancy outcome EP are shown per strategy, i.e. an absolute single serum hCG level, serum hCG ratios and logistic regression models, in ROC plots in Fig. 2. These curves show that an absolute single serum hCG level had a poor performance in diagnosing EP in women with PUL, compared with the serum hCG ratios and logistic regression models, as the sROC curve is considerably closer to the diagonal, indicating no diagnostic value.

Discussion

We performed a systematic review of various serum hCG-based strategies to diagnose the pregnancy outcome in women with PUL. For most pregnancy outcomes, the data were clinically very heterogeneous, and only the outcome EP was suitable for meta-analysis. An absolute single serum hCG level had the lowest diagnostic value, while strategies using serum hCG ratios, either alone or incorporated in logistic regression models, showed reasonable diagnostic performance for EP.

Worldwide, different diagnostic strategies using serum hCG in women with a PUL have been introduced and optimized within the same population for more accurate and earlier diagnosis. Differences in definitions of populations (e.g. very strict ultrasound definitions of PUL versus any form of inconclusive scanning as defined by the clinician) and final pregnancy outcomes (e.g. IUP solely based on location therefore a combination of viable and non-viable pregnancies, or very strictly defined as a pregnancy leading to the delivery of a child), have made the interpretation of current literature difficult and a meta-analysis of the data nearly impossible. It is important to realize that geographic differences in the prevalence of, for example EP and the a priori risk of the women enrolled, may affect the performance of a particular strategy. This is illustrated by the validation of a logistic regression model developed in the UK on a US population (Barnhart et al., 2010). Whereas this model had 80% sensitivity for EP in the UK, this decreased to 49% when applied in the US PUL population, with a similar specificity of ∼88% in both populations. A first step in improving the objective comparison of research outcomes in the diagnosis of PUL and in reducing clinical heterogeneity is the international consensus statement which has recently become available (Barnhart et al., 2011). This consensus statement proposes uniformity in definitions of population, target disease and final outcome of women with a PUL to allow the objective interpretation of research, objective assessment of future reproductive prognosis and hopefully lead to improved clinical care.

The results of the pooled data need to be interpreted with caution. First, all included studies were cohort studies evaluating one diagnostic test for the diagnosis EP. To date, no randomised controlled trials have been published comparing different diagnostic strategies for EP in women with PUL. Second, there was clinical heterogeneity between the studied populations. Most studies originating from the USA reported the evaluation of women who presented to emergency departments, where the ultrasound distinction between PUL and probable EP is not always made. In studies from European countries, women are directly evaluated by a specialized team in an early pregnancy unit, which may have resulted in a stricter ultrasound definition of PUL. Third, the reference test for the diagnosis EP was not uniform in all studies. The gold standard in the diagnosis of EP is laparoscopic confirmation but the PUL population is mostly asymptomatic and also a candidate for non-surgical management. Also, laparoscopy has some limitations; early ongoing EPs are sometimes too small to be seen at the time of laparoscopy. In addition to laparoscopy, alternative ultrasound reference tests for EP were used: an ectopic gestational sac with the visualization of a yolk sac or embryo, an inhomogeneous mass (blob sign) or an empty gestational sac (bagel sign). This alternative reference test using ultrasound for the diagnosis of EP is also not without limitations. However, the diagnostic performance of TVS can achieve a sensitivity of between 93.2 and 94.4% (Condous et al., 2005c; Kirk et al., 2008). Another reference test for EP was an increasing serum hCG after uterine curettage with absent chorionic villi (Barnhart et al., 2002).

In general, the pregnancy outcome EP alone is not the most patient-centred outcome. Owing to the earlier diagnosis of EP, the field is moving from the clinically important EP that requires surgery because of a real risk to woman's health, to those EPs where health practitioners act preventatively (e.g. with MTX) despite the fact that spontaneous resolution is possible. Current literature on the diagnostic accuracy of serum hCG strategies in women with PUL did not provide enough data to inform us about diagnostic criteria needed to distinguish between pregnancies that will lead to a viable IUP, those that do not require any intervention because the pregnancy is failing and those where we need to intervene as failure does put the woman's health at risk. Properly conducted prospective studies are needed to provide this relevant data.

Some diagnostic strategies in single studies had rather low sensitivity and/or specificity. The natural history of PUL may be such that more time is needed (i.e. testing later in the pregnancy) to distinguish the pregnancy outcomes. Rendering a prediction on only a single, or two, serum hCG measurement may lead to an inherent misclassification.

The sROC curves based on the pooled data show a higher sensitivity for the serum hCG ratio strategy but the regression models apparently have a better performance regarding specificity. In clinical practice these models are complex and often need computer analyses. Since most women with PUL are followed up by at least two serum hCG measurements, the serum hCG ratio strategy seems easier to implement.

In conclusion, in women with PUL the diagnostic strategies using serum hCG ratios, either alone or incorporated in logistic regression models, seem to have the best diagnostic performance in case of EP. Overall, the data were very heterogeneous and comparative diagnostic studies have not been performed on the diagnostic value of serum hCG in women with PUL. Studies directly comparing the efficacy of serum hCG strategies are warranted in order to identify the best strategy to diagnose pregnancy outcome in women with PUL.

Authors' roles

W.M.A., K.B., B.W.M. and F.v.d.V. all contributed substantially to the conception and design of this review. NvM and FM screened all titles, abstracts, articles and extracted data for meta-analyses. B.W.M. and P.J.H. were third reviewers in case consensus could not be reached directly. B.C.O. supervised the analysis and interpretation of data. N.v.M. drafted the article, all other authors critically revised multiple versions of the manuscript. All authors gave their final approval of the version to be published.

Funding

F.M. and P.J.H. are supported by grants from The Netherlands Organisation for Health Research and Development (Agiko stipendium 920-03-328 and Clinical Fellow 40-00703-97-05-154). No other funding was either sought or obtained for this study.

Conflict of interest

None declared.

Acknowledgements

We thank C.E.J.M. Limpens, clinical librarian, for her help with the literature search.

References