Incidence rates for AFE with 95% confidence intervals were calculated using as a denominator the number of maternities, deliveries or live births recorded regionally or nationally during each study period. Putative risk factors, identified from factors previously reported in the literature (maternal age, parity, smoking status, race/ethnicity, diabetes, multiple pregnancy, previous caesarean delivery, hypertensive disorders, placenta praevia, placental abruption, presentation at delivery, chorioamnionitis, polyhydramnios, induction of labour, mode of delivery, manual removal of placenta, macrosomia, gestational age at delivery) [1-4,22], were examined using univariable or full multivariable logistic regression analysis where possible and results presented as odds ratios (ORs) with 95% confidence intervals. Where individual level comparison group data were not available, associations were examined by comparing data on the proportion of cases with a particular risk factor with data on the proportion of the population of women giving birth with the same risk factor using risk ratios with 95% confidence intervals.

The UK AFE study was approved by the London Multi-centre Research Ethics Committee (ref 04/MRE02/46) and the linkage and use of New South Wales data was approved by the NSW Population and Health Services Research Ethics Committee (2006-06-011). The Dutch LEMMoN study was centrally approved by the medical ethics committee of the Leiden University Medical Centre (ref P04-020). No other permissions were required for use of the data presented.

The reported incidence of AFE ranged from 1.9 cases per 100 000 maternities (UK) to 6.1 per 100 000 maternities (Australia) (Table 2). There was a clear distinction between rates estimated using different methodologies. The lowest estimated incidence rates were obtained through validated prospective case identification (range 1.9-2.5 cases per 100 000 maternities); rates obtained from retrospective analysis of population discharge databases were significantly higher (range 5.5-6.1 per 100 000 admissions with delivery diagnosis), with an intermediate estimated incidence obtained from analysis of a regional discharge database in which additional required criteria (at least one of the cardinal symptoms with no other potential explanation) were used to exclude false positive cases (3.3 cases per 100 000 maternities). When the latter database was used without the additional selection criteria, the estimated incidence was 6.3/100,000.

Mortality ratios ranged from 0.4 per 100 000 live births in the Netherlands from 1993-2005 to 1.3 per 100 000 live births in the United States in 1997-2001 and 1.1 in Australia excluding Victoria in 1994-2005. Case-fatality rates ranged from 11 to 43%. There were no clear differences in estimates of maternal mortality due to AFE or case fatality rates due to AFE using any different methodology, noting the limited power to detect differences due to small numbers of fatal cases (Tables 3 and 4). The studies with the highest case fatality rates (NSW and Victoria) were conducted using case validation. Note, however, the wide confidence intervals surrounding all case fatality estimates due to the rarity of the condition and the small numbers involved.

The only factors consistently associated with AFE across all five countries were induction of labour and maternal age; although with both of these factors, in one country the association was not statistically significant. The association with age was not statistically significant in the Dutch data, this could be due to limited study power since all cases occurred in women who were aged 29 years or greater (Table 5). The methods of induction investigated varied. UK and Dutch data show a significant association between AFE and induction of labour (all methods); in the Canadian population, an association was observed with medical induction of labour, whereas in New South Wales there was only a statistically significant association with medical induction using vaginal prostaglandin. In the US study, induction of labour (all methods) was associated with raised odds (aOR 1.5), but this was not statistically significant. Where data were available, there were increased odds of AFE associated with placenta praevia and placental abruption. For all of the other factors examined (Tables 5, 6, 7, 8), the associations varied in the direction of effect and statistical significance, which may reflect small numbers and consequent lack of study power.

Other associations were noted with mode of delivery: forceps/vacuum and caesarean section, although these are challenging to interpret since information on the timing of delivery in relation to the AFE was not available for all of the data sources. In the UK data, where timing of the event and delivery was available, there was a statistically significant association noted with caesarean section delivery when the AFE occurred after delivery. There was no association with forceps or vacuum delivery, although the small number of women with AFE who had operative vaginal deliveries means there is limited statistical power to examine this association. Of note, eclampsia, a condition which may form one of the differential diagnoses for AFE, was strongly associated with AFE in the Canadian and US studies which did not use additional criteria to exclude false positive cases.

Very limited data were available on factors associated with fatality (Table 9); there were no factors consistently associated with fatality across all of the countries.

For the majority of sources used, there was very limited information on maternal outcomes other than death. Cerebral injury was noted in 6% of women with AFE in the UK, and cerebral infarction occurred in 20% of women with AFE in New South Wales. Data were similarly limited on fetal and infant outcomes. Eight of 21 infants (38%) born to mothers with AFE in the Netherlands were stillborn or died in the neonatal period; the figure was 5 of 75 infants (7%) in the UK and 6 of 19 (32%) in NSW.

This analysis demonstrates differences in the reported incidence of AFE in high-resource countries; the reported incidences vary according to the study methodology. Incidence estimates generated from analysis of population databases without additional criteria to exclude false positive cases produce more than double the estimates generated from analyses using specific validated case identification; an estimate from a population database analysis with additional criteria to exclude false positive cases was compatible with the estimates from the validated case identification studies. Comparison of the AFE incidence using a population database with and without criteria to exclude false positive cases also found this pattern, with the estimated incidence from the unselected population database double that when the additional criteria were used. This suggests that the noted European-North American differences in incidence [1] are also likely to be explained by the differences in methodology used, and may not represent a true difference in incidence between the countries.

It is difficult to determine which of these estimates is likely to represent a value closest to the "true" incidence of AFE. The apparently lower case fatality amongst cases identified through database studies supports the argument that these studies include a number of false positive cases. The observed association with conditions which may form part of the differential diagnosis of amniotic fluid embolism, for example eclampsia, also supports this hypothesis. Conversely, the higher case fatality observed in studies with case validation may simply indicate that these studies identify more severe cases. Database studies without case validation may thus have high sensitivity but conversely low positive predictive value due to the inclusion of a number of false positive cases, whereas studies with case validation may have lower sensitivity, due to lower case ascertainment and false negative cases, but higher positive predictive value. Perhaps the gold standard would be a database study with subsequent case validation through examination of the medical records. Such an approach, however, in the context of a rare condition such as AFE is difficult due to the large number of individual centres at which medical records would need to be accessed. All studies are complicated by the additional fact that AFE is a diagnosis of exclusion; internationally accepted diagnostic criteria for non-fatal cases do not exist, and there is a place for development of such criteria, using as a model, for example, the criteria used within the UK [3]. Differences in the case definitions used are likely to add to the observed variation in incidence estimates.

Thus the optimal approach to investigating AFE may be dependent on the purpose of the study, since each of the practical approaches has both advantages and disadvantages. Population-based database studies, where suitable sources of information exist, are likely to be the most economical option for studying the condition; however, there is likely to be a degree of inclusion of false positive cases, over-estimation of incidence and thus non-uniformity of the cases examined. The use of additional criteria to exclude false positive cases within database studies, the approach taken in New South Wales [22], may overcome some of these issues. However, without review of the medical records, this approach may exclude true cases that have been poorly coded. Database studies are also limited by a restricted number of available data items; according to the purpose of the original data collection, extremely limited information on management is available, and temporality of events, for example in relation to whether the AFE occurred before or after birth, cannot be studied. For the purposes of examining risk factors and management in detail, studies with specific prospective data collection methods may therefore be the best approach. Such studies are facilitated by available collaborations designed specifically to study rare pregnancy disorders, such as the approach taken by UKOSS [3], the Australasian Maternity Outcomes Surveillance System (AMOSS) and other members of the International Network of Obstetric Survey Systems (INOSS) [23].

Recommendation 1: Comparisons of AFE incidence estimates should be restricted to studies using similar methodology. Depending on the available resources and research questions, the recommended approaches would be either population-based database studies using additional criteria to exclude false positive cases, or tailored data collection using existing population-based systems specifically designed to facilitate study of rare pregnancy conditions.

Recommendation 2: Comparisons of AFE incidence between and within countries would be facilitated by development of an agreed case definition, particularly for non-fatal cases, and an agreed set of criteria to minimise inclusion of false positive cases for studies conducted using population-based databases.

We identified very little consistency among countries in the factors associated with occurrence of AFE and no factors consistently associated with fatality from AFE. Maternal age appeared to be associated with the occurrence of AFE in all populations examined, and this may therefore represent a true association, although the mechanisms by which older maternal age predisposes to AFE remain hypothetical and may include disruption of the uterine vasculature or minor degrees of abnormal placental invasion. The available information on placental abnormalities suggests that placenta praevia and placental abruption substantially increase (3-10 times) the risk of AFE. Given the differing methodologies and therefore differing biases due to potential inclusion of false positive cases or selection of more severe cases, other reported associations are very difficult to interpret. One of the most widely reported associations, with induction of labour, was noted to be statistically significant in some studies (UK, Netherlands), with only certain induction methods (Australia, Canada) or the association was not statistically significant, although the estimated odds ratio was compatible with those reported in most other studies (US) [2]. This illustrates the difficulties of comparing data across studies with differing case identification processes and definitions and varying levels of detail about individual risk factors.

With other factors, such as mode of delivery, issues concerning the timing of the event make the data difficult to interpret. The data we examined were often limited, frequently because the timing of the AFE in relation to the delivery was not available, and hence we were unable to identify cases in which a caesarean or operative vaginal delivery was a cause and not a consequence of the AFE. In the studies where timing was available, small numbers of women in certain subgroups, for example the number of cases undergoing operative vaginal delivery, limited statistical power. It is important to note, however, in the context of rising caesarean delivery rates worldwide, that there was a significant association with caesarean delivery in the one country (UK) in which we were able to investigate specifically cases where the AFE occurred after delivery. Analysis of pooled international data, obtained using consistent methodologies with agreed definitions and case validation, would provide more reliable information on these associated factors and hence provide the potential to develop appropriate preventive strategies which would otherwise be limited.

Recommendation 3: Groups conducting detailed population-based studies on AFE should develop an agreed strategy to allow combined analysis of data obtained using consistent methodologies in order to identify potentially modifiable risk factors.

There were no clear differences in maternal mortality ratios due to AFE between countries or by study methodology. Information on other outcomes for women was very limited, although data on cerebral injury [3] suggest that AFE may have a significant impact in the long term outcomes of survivors, and this requires further investigation. None of the data sets included information on long-term maternal outcome after non-fatal AFE. Data on infant outcomes were only available for sources which can link data on the mother and infant or from studies which collect specific data. Although limited, the available data suggest the perinatal death rate associated with AFE is high (7-38%). Similarly, we were unable to investigate any potential relationship between management and outcomes because of extremely limited data. Further international collaborative studies using specific data collection would allow for more detailed investigation of the outcomes for mother and infant, over both short and longer term, and any relationship with management.

Recommendation 4: Future specific studies on AFE should aim to collect information on management and longer-term outcomes for both mothers and infants in order to guide best practice, counselling and service planning.

This analysis used data from five high resource countries, and the results are thus generalizable only to countries with similar resource settings. As discussed above, data on AFE were obtained using different study methodologies, and therefore this limits the comparability of some of the results. AFE is a rare condition, and all studies have limited power to detect true associations as statistically significant. We hope that this may be addressed in the future by further international collaborative studies, an approach that we would advocate for research into all rare conditions in pregnancy.