Gestational diabetes mellitus in relation to serum per- and polyfluoroalkyl substances: A scoping review to evaluate the need for a new systematic review

Introduction

Per- and polyfluoroalkyl substances (PFAS) were used or have ongoing use in the manufacturing of a variety of products, including consumer-use products.^1,2 Almost everyone in developed countries has a detectable amount of PFAS in their blood serum due to exposure via contaminated food and other sources, and the unreactivity of these chemicals.³ Despite some being phased out of production,⁴ PFAS are widespread and have been detected in different continents irrespective of the level of industrialization, indicating long-range atmospheric transport as an important pathway of PFAS distribution.⁵

Serum or plasma concentrations of PFAS have been associated with several health outcomes, including gestational diabetes mellitus (GDM).^6–11 GDM increases risk among offspring of preterm birth and macrosomia,¹² and, among mothers, of type 2 diabetes mellitus.¹³ In the U.S., the proportion of pregnant women who develop GDM has increased over the past ten years, possibly due to trends in risk factors such as obesity, but environmental factors could be contributing.¹⁴

Our recent experience in evaluating systematic reviews of the relation of health outcomes to PFAS serum concentrations has made us aware of the variety of comparators used in original reports.¹⁵ This variety poses challenges in summarizing data that have not been adequately addressed in the GDM-PFAS systematic review literature. Thus, in addition to identifying relevant original reports in this scoping review, we addressed our special interest in the methods used in previous meta-analyses – but not in their findings.

Methods

We followed Joanna Briggs Institute’s guidance for designing and conducting this scoping review¹⁸ and followed PRISMA-P,¹⁹ PRISMA-ScR,²⁰ PRISMA-Abstract,²¹ and PRISMA-Search²² in reporting the protocol and the final report of this review. The indications for a scoping review were: a) to examine how research is conducted on a certain topic or field, and b) as a precursor to a systematic review.²³

We used the PECO framework (Population, Exposure, Comparator, and Outcomes)²⁴ for formulating the problem to be addressed. The PECO also guided setting the eligibility criteria (Table 1). The population used was pregnant women. The exposure was the concentration of any specific PFAS in serum or plasma during pregnancy. The measure of exposure was obtained from a blood specimen or, for populations in which a validated pharmacokinetic model of serum concentration was available, estimated on an individual basis. Estimated serum or plasma PFAS for an individual from a population for which a validated pharmacokinetic model of serum concentration was available is acceptable because in the absence of a serum or plasma measurement, pharmacokinetic models can be useful for imputation when a population water supply had unusual high PFAS concentrations.²⁵ The comparator was a contrast in serum PFAS concentration that was either per arithmetic unit of PFAS (e.g., ng/ml), per log unit increase, or categorical. The primary outcome was GDM, and the secondary outcome was serum glucose.

Observational studies such as case-control, longitudinal (cohort, prospective, retrospective), or cross-sectional studies were included in this review. Full reports about original epidemiologic research or meta-analyses had to be published online or in print. Pre-clinical studies (cells or genetic material) and studies with non-human subjects were excluded. If the data were for a non-pregnant population only or a mixed population where separate data on the pregnant population was unavailable, we excluded the report. If a measured or estimated serum or plasma concentration of PFAS in the person was not examined in relation to the occurrence of GDM or blood glucose in the person the report was excluded. Mixture analyses where there were no single exposure analyses for PFAS reported in that paper were excluded.²⁶ We excluded narrative and systematic narrative reviews; however, before the exclusion, we checked for any relevant unique included reports in the systematic narrative reviews. To identify relevant ongoing or published systematic reviews to examine their methods, we conducted a rapid search in PROSPERO and PubMed.

On 11^th December 2022, we systematically searched the latest available version of the following sources to identify original reports without date, time, document type, or language limitations²⁷: CINAHL via EBSCOhost (1937 – search date); Embase via Ovid SP (1974 – 9^th December 2022); Emerging Sources Citation Index via Web of Science (2015 – search date); MEDLINE via Ovid SP (1946 – 9^th December 2022); PubMed (excluding MEDLINE; 1946 – search date); and Science Citation Index Expanded via Web of Science (1900 – search date).

An information scientist designed the search strategies using the terms collected from the experts, literature, and controlled vocabularies (CINAHL Headings, Medical Subject Headings=MeSH, and Emtree). The search strategy for MEDLINE was shared with two experts on the topic for commenting and revisions and peer-reviewed by another information specialist based on PRESS guidance²⁸ before being translated into the syntax of other databases.

The search strategies (See the Extended data) were tested and tuned based on a randomly selected half of the studies included in the previous systematic reviews. The final search was tested against the remaining half of the included studies from previous reviews to ensure the comprehensiveness of the retrieval.

Results were imported to the reference management software, EndNote X9, for de-duplication. The de-duplicated search results were imported into Rayyan²⁹ and two reviewers screened the search results independently after turning the blinding mode on. We followed the same procedure for screening the full reports of the studies. For any disagreement, the two reviewers discussed and agreed on a final decision to include or exclude the report.

Data extraction was designed, piloted, and revised to include the following data points: study name (first author’s last name and publication year), study design, study period, country of origin, participants’ age, number of participants per group, duration of the pregnancy at the baseline of the study, time of glucose measure, name of the glucose test, diagnostic criteria for GDM, sample type for exposure measurement, time of measuring PFAS, ethnicity, setting (rural, urban), source of exposure, length of exposure, name of substances, summary statistic for PFAS and glucose concentrations per group, exposure data type, chemical analysis methods, measures of association, exposure metric (comparator), potentially relevant references for identifying more relevant studies, the lead author’s name and email address, and any relevant notes.

When we could not find the required data in the report, we contacted the authors to request it.

We summarized the findings narratively or in tables. Because this was a scoping review, we did not run any data analysis or risk of bias appraisal for the included studies.

Results

We identified three previously published systematic reviews that explored the GDM-PFAS association (Table 2).^30–32 In one systematic review (Yan et al., 2022),³² the use of meta-analysis to combine results across comparators rendered the meaning of the summary OR obscure. In another systematic review (Gao et al., 2021),³⁰ results from different evidence streams were not combined using an appropriate summary. In the third systematic review (Wang et al., 2022), the focus on only high:low exposure category comparators meant: a) informative data from each included study were ignored, b) the variability of the exposure contrasts made use of this comparator questionable, and c) reports using only models with exposure as a continuous variable were ignored. The most recent of the systematic reviews had a literature search end date of November 2021. None of the previous systematic reviews employed synthesis without meta-analysis (SWiM).³³ Although our focus was not on the results of the previous systematic reviews, a summary of their results was as follows: a) no significant association was found between PFAS and GDM³⁰; b) PFOA was significantly associated with higher risk of GDM³¹; and c) PFOA and PFBS increase risk of GDM.³²

Four protocols registered in PROSPERO either explicitly addressed GDM and PFAS or might have included it in a broader definition of outcome or exposure (Table 3). Two of the protocols were old enough that it was likely the effort had been abandoned; of the more recent protocols, one did not specifically mention the inclusion of pregnant women or gestational diabetes. The remaining study described the use of a high:low exposure category comparator, which is statistically inefficient because it does not include all available results from a report and has other limitations (see Discussion). None of the studies mentioned use of SWiM.³³

The search for original data reports found 4938 records. After removing the duplicates, we screened the unique records and identified 82 possibly relevant records based on their titles and abstracts. When the full-text reports were obtained for these 82 records, we excluded 62 after documenting the reason for exclusion and included 20 reports in this review.^{6–11,34–47} The process of selection of reports is shown in Figure 1.

Figure 1. The PRISMA diagram for the process of selecting the studies.

Two reports used the data from the Shanghai Birth Cohort^11,45 with the possibility of duplicate and overlapping data. The source and length of exposure were not reported in any of the reports, presumably indicating that all were of populations with background exposure.

Characteristics of the reports are presented in Table 4. Based on these data, the prospective cohort (12 reports) was the most used study design followed by case-control (5). The studies were conducted between 1997 and 2021 in China (10 reports), North America (US (6) & Canada (1)), Denmark (2), and Spain (1). The 10 studies from China all had Chinese participants, the ethnicity of participants from all seven North American studies was diverse, one of the studies from Denmark included only participants of Danish ethnicity, while the other did not report participant ethnicity, and the one study from Spain had participants of primarily Spanish origin.

The 20 reports included a total of 18,805 participants. Among the 15 reports that presented the number of participants with GDM, the proportion with GDM was 10% (1,655/16,513); the other five reports presented results on serum glucose only and not GDM. The number of participants across the reports ranged between 95 and 2747. The participants’ age was between 20 and 40 with the most frequent mean age reported being between 28 and 32.

Participants’ duration of pregnancy at study baseline varied between 5 and 38 weeks making it one of the most heterogenous items of data extracted.

Table 5 summarizes the list of 24 PFAS (not counting isomers) for which a measure of association was reported in the included studies with results for PFOA, PFOS, PFHxS, PFNA, and PFDA being reported in almost all the studies.

Table 6 presents the data regarding the measurement and reporting of PFAS concentration, glucose, and GDM. An oral glucose tolerance test (OGTT) was used in 16 reports as the glucose test and seven reports followed the International Association of Diabetes and Pregnancy Study Groups (IADPSG) criteria for diagnosing GDM. Among 12 reports that reported type of glucose measure, 10 reported fasting glucose (FG), six reported one hour glucose (1hG), and five reported two hour glucose (2hG), most of which reported mean ± standard deviation (nine reports). While the time for measuring glucose levels varied between 14 weeks of gestation and 1-2 days pre-delivery, measurement of PFAS was in a wider time range from 8 weeks of gestation to delivery. HPLC LC-MS/MS was the most-used chemical analysis method for PFAS. A summary statistic for PFAS concentrations was presented for 19 reports with the geometric mean being the most frequently reported (10) followed by the median (seven reports). Standard deviation was presented in only four reports with geometric means and interquartile range (IQR) for six reports with median.

To measure the association between PFAS concentration and glucose, the reports used associations between PFAS and one or more of the glucose measures: FG, 1hG, 2hG, OGTT (average of 1hG and 2hG), Impaired Glucose Tolerance (IGT), and GDM. Adjusted odds ratio (OR) (12 reports), crude OR (eight), adjusted regression coefficient (β) relating serum glucose to PFAS (nine), and crude β (three) were the main reported association measures. When a log exposure metric (comparator) was used, it was implemented heterogeneously with per one log-unit difference in PFAS concentration presented in nine studies, per log2 (doubling) difference in four, and per log10 difference in three. Eleven of the reports examined associations with exposure data categorized as either tertiles (six reports), quartiles (four), or dichotomized (one). The majority of these were in conjunction with exposure data examined as a continuous variable, but two reports displayed results with categorized PFAS concentration only, one as tertiles and one as quartiles of exposure.

We summarized the type of data in the original reports identified by our literature review according to the comparator and outcome (Table 7). Results from reports that reported an effect measure for GDM (e.g., odds ratio) with exposure categorized into more than two categories were counted in the table as being arithmetic comparators because the results can be re-expressed that way.⁴⁸ For GDM, we found that 12 reports could be combined in a SWiM analysis for the two major PFAS. For reports presenting effect measures for glucose during pregnancy (same two PFAS), we found 11 results that could potentially be combined using SWiM. If results for GDM and glucose were combined in a SWiM, results for 17 could be included – assuming the participants in Wang et al. 2023a were a subset of those in Yu et al. 2021.

Discussion

This scoping review of the evidence and methods used in primary and secondary data on the relation of GDM with PFAS addressed the overall question of whether a new systematic review was needed. Because data from several new reports were available since previous systematic reviews were conducted and because SWiM³³ could make the best use of all available data, an updated systematic review using SWiM would be a valuable addition to the literature.

The variety of comparators used in the original reports poses challenges to the use of meta-analysis to summarize data. If the original reports include a categorical exposure scale, a meta-analysis using a high:low exposure category comparator has been presented as an option.²⁴ The weaknesses of this approach, however, merit discussion. For PFAS, the exposure contrast corresponding to the high:low comparison is likely to vary substantially across studies. If a dose-response relation is operative, the magnitude of the meta-analytic summary would reflect the distribution of exposure in the studies, and the validity of the resulting meta-analytic statistics associated with such a summary is questionable.⁴⁹ Furthermore, categorical results within-study can be re-expressed to an arithmetic effect measure, which solves the above-mentioned problem and has the further benefit of using, rather than a portion, all the available data. Another issue is that if arithmetic and logged exposure measures were interchangeable after some type of re-expression, that would facilitate meta-analysis. But such re-expression does not work reliably.⁵⁰ An interesting example of the underlying problem can be found in a recent report by Padula et al. (2023).⁵¹ They analyzed individual-level data on birth weight and serum PFOA first using log (exposure) and again using arithmetic exposure. The association was inverse in both analyses, but it was statistically significant when log (exposure) was used but not when arithmetic exposure was used. The SWiM is the only widely-accepted non-narrative approach to synthesizing the data from evidence streams based on different comparators.³³ The three previous systematic reviews each had methodologic elements yielding questionable results. Specifically, a) Gao et al.³⁰ did not provide an overall summary of results from different evidence streams, b) Wang et al.,³¹ by using a high:low comparator, ignored part of the data and combined results in the face of varying exposure contrasts, and c) Yan et al.,³² combined results across comparators, rendering nebulous the statistics associated with the results.

Another potential improvement in a future systematic review of the GDM-PFAS association would be to integrate data from the GDM-PFAS and blood glucose-PFAS evidence streams. Among studies of blood glucose as the only outcome (no GDM analysis), four out of the five included participants with GDM. Because blood glucose is near to normally distributed, the odds ratio for GDM and the beta coefficient for blood glucose are, with re-expression, interchangeable metrics.⁵² Thus, data from the glucose-PFAS reports could be used to address the overall question. Even without such re-expression, SWiM could be used to combine the evidence streams.

Results for five PFAS (PFOA, PFOS, PFHxS, PFNA, and PFDA) were reported in almost all the original data studies because these are the PFAS that have the highest median serum concentration in background-exposed populations. In the previous systematic reviews, Gao et al. (2021) reported results for all five³⁰; Wang et al. (2022) reported on four,³¹ and Yan et al. (2022) reported on these five plus three additional PFAS.³² PFAS with lower median concentrations could have been analyzed by the original authors with the comparator > LOD vs not, and such results have not, to date, been included in meta-analyses; however, they could be included in a SWiM.

When conducting a meta-analysis of odds ratios for GDM, an assumption is that the log of the odds ratio per comparator unit (beta coefficient) does not vary by the prevalence of GDM in the underlying populations. The prevalence of GDM in the underlying populations is a function of the diagnostic criteria, the prevalence of obesity, and other study or population characteristics.^53–55 Whether these characteristics modify the beta coefficient could be evaluated in a meta-regression; such regression was not done in the previous meta-analyses.

Time has passed since the date of the search for relevant original data reports. Our search for relevant systematic reviews, however, was much more recent, and it is this search that is most important for our conclusions to remain valid. The number of new reports available was not as important for our conclusions as the finding that the methods employed in previous systematic reviews could be substantially improved.