Changes in concentrations of cervicovaginal immune mediators across the menstrual cycle: a systematic review and meta-analysis of individual patient data

We included studies reporting original data on any immune mediator concentrations by menstrual cycle phase (determined by date of last menstrual period [LMP] or hormone levels, including progesterone, estradiol, and/or luteinizing hormone) in CVT samples from menstruating women. Immune mediators were defined as immune-related proteins, including cytokines, chemokines, immunoglobulins, antimicrobial peptides, and growth factors. We only included studies that measured concentrations using antibody-based methods (such as ELISAs, Luminex and other bead-based assays, and MSD assays). We did not include studies using other methods, such as gene expression or mass spectrometry-based proteomics or metabolomics. CVT samples were defined as secretions or fluid, such as CVL, menstrual cup, or swab. We included unpublished studies that met our eligibility criteria.

Participant-level eligibility criteria allowed us to include subsets of participants from studies where only some subjects were eligible (such as studies comparing pre- and post-menopausal women, where only the pre-menopausal women were included). Eligible participants were post-menarche, pre-menopausal, non-pregnant women not using hormonal contraception or an intrauterine device (IUD) and not receiving other exogenous hormones. Because intra-study comparisons of follicular and luteal phases were performed, each study had to have both follicular and luteal phase samples, but single samples from individual participants were eligible. We excluded participants who received a vaginal intervention (including placebo), but participants receiving no treatment or a systemic placebo were eligible. Baseline, pre-intervention visits were acceptable (such as if all participants had baseline visits, a cross-sectional analysis could be performed). Samples from women with cervical or vaginal pathology, such as bacterial vaginosis, vulvovaginal candidiasis, STIs, or cervical dysplasia, were eligible. We chose to include such samples because cervical or vaginal pathology is a normal part of life for most women at some point. In addition, we expected pathology to have no association with cycle phase (for example, we expected BV to be equally common in both phases of the cycle), so it would not confound our menstrual cycle analysis.

Search results were de-duplicated using PubMed IDs, the text of the titles and abstracts, and manual review of duplicate DOIs. Abstracts were loaded into abstrackr [37] for screening. Two reviewers (CNL and SMH) independently screened all abstracts for eligibility. We obtained the full text of all articles identified as potentially eligible by either reviewer. Both reviewers independently reviewed full texts, guided by a Google Forms questionnaire (Additional file 1) to determine eligibility and record study information. We recorded reasons for exclusion of a study in the questionnaire. Differences in opinion were resolved by discussion. If the two reviewers were unable to agree, a third study author (FH) made the final decision. If conference abstracts appeared to meet inclusion criteria, but could not be linked to a publication, we contacted the authors to locate the publication. We attempted to extract summary data from all studies using the Google Forms questionnaire (Additional file 1). Specifically, if available, we extracted estimates of the difference in concentrations of each immune mediator between the follicular and luteal phases, as well as the statistical methodology used to generate that estimate. We anticipated that this summary data would be unavailable from many manuscripts.

We requested individual participant data (IPD) from study authors via email, following up at least three times. We accepted data in any format provided. After receipt of IPD, we prepared a data summary document (including the number of samples, number of immune mediators, menstrual phase, covariate summaries, the number of samples below LOD, and the immune mediator means and 95% CIs). We sent this summary document to the study authors and requested that they confirm that we received the complete and correct data. We also compared the IPD we received and the results of our analyses to published reports, where available, to confirm that the data we received was correct.

If we were unable to obtain IPD for a particular study, we recorded the reasons that prevented obtaining the data and attempted to extract IPD from the published article. Two reviewers independently extracted the data and discussed differences, with a third reviewer resolving discrepant results and disagreements when necessary. Data were extracted from published figures using software such as WebPlotDigitizer [38], if appropriate.

If IPD was unavailable from the authors and could not be extracted from the published article, we recorded the reasons that prevented obtaining the data. If summary data was available (differences between follicular and luteal phases, extracted for all studies as described above) and matched the study-level analyses described below, we included the study at the meta-analysis level in the two-stage approach described below. For papers where only quantile statistics were reported, we obtained means and standard deviations (necessary for meta-analysis) using previously devised methods [39‐42].

We collected the following study-level data items:

We collected the following sample-level data items:

We collected the following immune mediator-level data items:

The definition of menstrual phase was standardized across studies and based on either serum progesterone level, days since luteinizing hormone (LH) surge, or days since the start of the last menstrual period (LMP). If multiple measures were available, we defined the menstrual phase based on hormone levels. For serum progesterone, the follicular phase was defined as serum progesterone < 1 ng/mL, and luteal was defined as serum progesterone ≥ 3 ng/mL. We chose these criteria based on a study [43] showing that the vast majority of pre-ovulatory samples have progesterone levels below 1 ng/mL and the vast majority of post-ovulatory samples have progesterone levels above 3 ng/mL. We excluded samples falling in the 1–3 ng/mL window, because these typically occur beginning on the day of the luteinizing hormone peak and ending two days after. For studies reporting LH surge without progesterone levels, follicular was defined as after menses and prior to LH surge, while luteal was defined as 2–12 days following LH surge. For studies reporting LMP, we only included participants reporting regular menstrual cycles. Follicular phase included days 5–12 (inclusive) since the start of the last menstrual period, and luteal phase included days 19–24 (inclusive) since the start of the last menstrual period. In some circumstances, decisions about sample inclusion were made on a case-by-case basis by discussion between two reviewers. The circumstances could include (1) samples falling outside the windows for days since the last menstrual period, LH surge, or progesterone concentration; (2) studies where hormone concentrations or days since LMP were used to determine menstrual phase, but those data are no longer available; or (3) studies where menstrual phase was determined by another method, such as urinary progesterone metabolite concentration.

We included periovulatory samples as a third phase, with this phase defined by LH levels above 20 mIU/mL in serum [44] or 25 mIU/mL in urine [45].

All additional variables were standardized across studies to the extent possible, based on the data. We defined assay type, sample type, and method of determination of menstrual phase as described above. We treated swabs from different anatomic sites (ectocervical, endocervical, vaginal) as different sample types. CVLs were considered a single sample type, but differences in methods of collection were explored in sensitivity analysis as described below. We assigned consistent cross-study definitions to additional covariates as much as possible based on the data collected. For example, for bacterial vaginosis (BV), if one study reported Nugent scores and another study reported BV based on Amsel criteria, we converted these variables into a single variable for BV, with values of positive, indeterminate, and negative.

If the limits of detection were unavailable, we attempted to obtain the information from the manufacturer of the assay. If the limits of detection were not available from the manufacturer, we classified the values as follows: undetectable when two or more samples have the lowest reported concentration for a given immune mediator in a particular study. Otherwise, samples were classified as detectable.

We assessed the risk of bias in each study using a custom tool adapted from the Newcastle Ottawa scale (Additional file 1). This information was used in determining the strength of evidence.

We performed meta-analysis for all immune mediators present in at least two included studies. Data analysis was performed using R version 4.0.0.

Data processing: Sample wells falling below the lower limit of detection were assigned a value of the study-specific lower limit of detection divided by 2. Wells falling above the upper limit of detection were assigned a value of the study-specific upper limit of detection multiplied by 2. If replicate wells were run for a given sample, the raw concentrations were averaged. Data was then log2-transformed. Each sample was also scored as “detectable” or “non-detectable”, with the sample counting as detectable if it was detected in at least one well.

We used a two-stage approach for meta-analysis: first analyzing each study separately and then combining the summary statistics from each study to generate meta-estimates of effect. We chose this approach to allow inclusion of studies where summary data was available but IPD was not.

For assay type and sample type, we performed meta-regression after the two-stage approach described above. In addition, we performed subgroup analysis stratifying by each covariate (assay type, sample type) and compared the standard error of the menstrual cycle effect sizes.

For the method of the menstrual phase, we analyzed studies that reported both hormone levels and days since the first day of LMP. For those studies, we performed the menstrual cycle analysis separately using each method of determining the menstrual phase. We then compared the standard errors within study.

For normalization to total protein, we only used data from studies reporting total protein concentrations. We performed the menstrual cycle analysis separately on the raw immune mediator concentrations and on the immune mediator concentrations normalized to total protein. We then compared the standard errors within each study.

We reported χ² tests and the I² statistic to summarize between-study heterogeneity in the menstrual cycle effect. For immune mediators with high levels of heterogeneity (I² > 75%), we attempted to explain the heterogeneity through subgroup or sensitivity analysis.

Sensitivity Analyses: The goal of the sensitivity analyses was to determine how robust the results were to analytic assumptions. We compared the results of several alternative analyses to the primary analysis described above.

Immune mediators measured in only one study or that could not be included in the meta-analysis for any other reason were listed as areas for further research.

We assessed publication bias and selective outcome reporting. We attempted to limit bias due to selective outcome reporting by requesting IPD for all immune mediators measured, regardless of which were reported in published studies. To attempt to limit publication bias, we sought out unpublished studies by requesting them from authors who contributed IPD from published studies and by including conference abstracts in our search strategy. To assess publication bias, we reported Egger’s test and funnel plots for immune mediators where ten or more studies existed.

If any issues with study data were uncovered when we checked the IPD, we reported these issues and any corrective actions taken.

We assessed the strength of the body of evidence using the GRADE methodology [46], with the instrument shown in Additional file 1. Two reviewers (CNL and SMH) performed the assessments independently and then came to a consensus, with disagreements resolved by a third author (FH).

We assessed the strength of the body of evidence for each immune mediator in five domains (risk of bias, inconsistency, indirectness, imprecision, and publication bias), each of which could lead to downgrading of the strength of evidence. We also assessed domains which could lead to upgrading of the strength of evidence, including large magnitude of effect (defining large as 5-fold and very large as 10-fold) and residual confounding that would be likely to strengthen the observed effect (or lack thereof). Randomization and dose responses were not be taken into account as they are not relevant for these studies (participants cannot be randomized to a particular phase of the cycle and dose is irrelevant for the cycle).

We assigned an overall strength of evidence score to each immune mediator based on a four-star scale as follows: high (further research is unlikely to change our confidence or the estimate of the effect), moderate (further research may change our confidence and the estimate of the effect), low (further research will likely change our confidence and the estimate of the effect), and very low (further research will very likely change our confidence and the estimate of the effect).

As part of this review and meta-analysis, we performed one additional study including an exploratory and a validation component. We used CVL samples from the Kenya Girls Study, a longitudinal cohort study of adolescent girls followed for acquisition of sexually transmitted infections [47]. We chose samples using the following requirements: no use of hormonal contraception, at least one follicular and one luteal phase sample available from the same participant (based on the date of LMP), STI testing and Nugent scoring for BV performed, and non-intermediate vaginal flora (Nugent score either 0–3 or 7–10). We measured serum progesterone to assign samples to the follicular or luteal phase. We measured total protein concentrations in CVL samples. Because sexual activity and exposure to semen may affect CVT immunity, we measured kallikrein-3 (also known as prostate-specific antigen). Similarly, blood contamination of the samples may influence immune mediator concentrations, so we measured hemoglobin. The sample size was designed to be approximately 200 samples from approximately 100 women. This size was determined based on feasibility and cost. All participants provided written, informed consent in the Kenya Girls Study as described in the main manuscript for that study [47]. Only deidentified samples were used as part of this study.

The purpose of the exploratory component of the study was to increase the strength of evidence for immune mediators that were measured in only few studies. We selected the mediators to be measured after we obtained data from all studies for meta-analysis. We chose approximately ten immune mediators that were measured in only 1–2 studies, with the total number of immune mediators determined based on cost and feasibility. We gave preference to mediators of particular biological interest based on the literature and preliminary results of the meta-analysis. We incorporated the measurements from the exploratory study into the final meta-analysis as an additional study.

Concentrations of selected immune mediators were measured using Meso Scale Discovery (MSD) R-Plex/U-Plex kits and ELISA. MSD assays were used where available because they allow simultaneous detection of multiple immune mediators in the same well. ELISA was used for immune mediators that were unavailable or cost-prohibitive by MSD. When ELISAs were used, they were purchased from R&D Systems wherever possible. To measure kallikrein-3, we used the Human Kallikrein 3/PSA DuoSet ELISA (R&D Systems, catalog DY1344). To measure progesterone, we used the Progesterone ELISA kit (Enzo Life Sciences, catalog ADI-901-011). We planned to measure hemoglobin in an MSD panel with other immune mediators, if compatible, or by Hemastix Blood ID Reagent Strips (Siemens).

Prior to running all of the samples, we chose the appropriate dilution for each analyte by running a pilot set of samples run with no dilution, 1:10 dilution, and 1:100 dilution (greater dilutions performed as needed). The diluent for CVL samples was 1% bovine serum albumin in phosphate buffered saline, unless a different diluent was required for a particular kit. The diluent for serum samples was the assay buffer provided with the Progesterone ELISA kit. We chose the dilution for each analyte that resulted in the largest proportion of tested samples in the detectable range.

MSD and ELISA were performed according to the protocols provided by the manufacturers. To limit batch/plate effects, we ran all samples from a given donor on the same plate, and we distributed follicular and luteal phase samples across plates.

The MSD data was analyzed using MSD Discovery Workbench software using the built-in concentration interpolation (typically four-parameter polynomial curve) and the concentrations were exported. For ELISA, concentrations were determined using a four-parameter polynomial curve. We analyzed the data from the exploratory and validation components of the study using the same two-stage process as for the studies collected from the literature, as described above. As for all other studies, the primary analysis was unadjusted, and in sensitivity analysis, we adjusted for covariates including hemoglobin, recent sexual contact, STI, and BV status.

We assessed whether there was evidence of non-publication of results (i.e., publication bias) for all immune mediators that were measured in at least 10 studies. The risk of publication bias was assessed using Egger’s tests and funnel plots, where asymmetry would be suggestive of possible publication bias (Fig. S1). There was no evidence of publication bias for any of these immune mediators.

The overall risk of bias at the study level was assessed using the instrument in Additional file 1. The risk of bias was generally low in these studies, as shown in the last column of Table 2.

We used the GRADE framework to assess the quality of evidence for all immune mediators as described in the methods. The GRADE ratings are listed in Tables 3 and 4 and Fig. 3. Overall, the evidence strength was high for 26 immune mediators, moderate for 12, low for 9, and very low for 6.

Based on this interim meta-analysis, we met our pre-registered statistical power threshold of 0.9 for one immune mediator, total IgG (power = 0.96). Therefore, we only performed a validation experiment for a single immune mediator, rather than 2–3 as specified in the protocol. We predicted that IgG would be lower in the luteal phase. We measured IgG by MSD in 200 CVL samples from 100 participants from Kenya (Fig. 4A), with a final sample size of 178 CVL samples from 99 participants after excluding samples with insufficient volume or where serum progesterone levels fell outside the limits of our menstrual phase definitions. We found that IgG was 0.342 log2 units lower in the luteal phase than the follicular phase, with p = 0.183 (Fig. 4B), so the direction of effect was as predicted, but the p-value did not meet our specified threshold for statistical significance of 0.05.

We measured two additional validation immune mediators despite not meeting the pre-registered threshold for power. We felt that the experiments had the potential to be instructive and would at minimum contribute additional data to the meta-analysis. We chose the two immune mediators with the highest estimated statistical power other than total IgG: CCL2 (expected to be lower luteal; Fig. 4A) and IL-1α (expected to be higher luteal; Fig. 4A). We measured each by MSD and confirmed CCL2 to be lower in the luteal phase (−1.36 log2 units, p = 9.4E−7) and IL-1α to be higher in the luteal phase (0.73 log2 units, p = 8.0E−4; Fig. 4B).

We used these same samples for exploratory experiments of immune mediators that were measured in few studies. We chose the following immune mediators (all measured in 1–2 studies at the time the reagents were ordered): MMP1, MMP7, CCL11, CD40L, IL-15, IL-16, and IgM (all by MSD), as well as GNLY and CTSD by ELISA. We also measured IgA, even though it did not meet our criteria for validation (power >0.9) or exploratory (measured in 1–2 studies) experiments; we included it because it was included in the multiplex IgA, IgG, and IgM MSD kit. As described in the methods, we also measured total protein concentrations by BCA assay, PSA levels by ELISA, and hemoglobin A by MSD. Measurements were available from 175 to 182 samples from 98 to 99 participants per immune mediator after excluding samples as described above or that failed QC. Concentrations of these immune mediators are shown in Fig. 4C. IL-15 was detected in fewer than 50% of samples, so it was analyzed using logistic models. All of these immune mediators were lower in the luteal phase than the follicular phase, except for CTSD (Fig. 4D). The data from this experiment is included in the main meta-analysis in Fig. 3, substantially increasing the number of samples as well as the list of immune mediators included in the final meta-analysis.

As a pre-specified sensitivity analysis, we performed a one-stage meta-analysis. Specifically, we pooled the raw data from all studies and assessed the effect of the menstrual phase in a single model per immune mediator, with participant and study as random effects. This approach differs from our primary analysis reported above, where we used a two-stage approach, first analyzing each study separately and then combining the results by meta-analysis. The results of this one-stage meta-analysis confirmed the results of our primary analysis (Fig. 6A, Pearson correlation coefficient r = 0.93 for correlation of effect sizes between one- and two-stage analyses).

Because different covariates were measured in each study, our primary analysis did not adjust for covariates. To test whether the observed differences in immune mediator concentrations between phases were affected by covariates, we re-analyzed each study, adjusting for all relevant covariates for each study. The exact covariates adjusted for in each study are listed in Table S4. The most common covariates were bacterial vaginosis and detection of red blood cells (RBCs). Several studies were omitted, either because no covariates were reported or because there were too few samples to perform multivariate analysis. In addition, many samples had to be omitted due to missing covariate information. Because some samples had to be omitted in the multivariate analysis, we repeated our univariate meta-analysis on just the samples that could be included in the multivariate analysis, to allow for direct comparison. Thus, the univariate meta-analysis reported in this section differs slightly from the primary analysis, due to the smaller sample size used here. The meta-estimates of effect size were highly correlated between the univariate and multivariate analyses (Fig. 6B; Pearson r = 0.82), confirming our primary results. However, the covariates measured in each study were highly variable and the sample size per study was often limited.

We noticed that one covariate in particular was associated with cycle phase: presence of RBCs or hemoglobin in the samples (Fig. 6C). Therefore, we assessed this covariate further in an exploratory analysis that was not preplanned. Six studies used methods that could detect microscopic levels of blood (hemastix, hemoglobin A MSD assay, or RBC counts), and three used visual inspection. Microscopic levels of RBCs were detected in more than half of the samples. In contrast, visual inspection classified few samples as containing blood. Across all methods, there was a consistent pattern of greater RBC detection in follicular phase samples.

As a secondary outcome, we wished to determine whether one type of sample yielded higher concentrations and detection rates for immune mediators (regardless of menstrual phase). Thus, we compared the immune mediator concentrations detected by menstrual cup, sponge, and swab to CVL (which was by far the most common sample type). For this analysis, we included all immune mediators that were measured in at least two sample types and where each sample type was used in at least two studies.

As shown in Fig. 7A, menstrual cup, sponge, and swab consistently resulted in higher total concentrations than CVL, as expected. For all three sample types, the concentrations were higher than CVL for every immune mediator (p<0.05 for 12/20 immune mediators by menstrual cup, 4/4 by sponge, and 0/6 by swab). The study-level concentrations are illustrated for one representative immune mediator (CXCL8, selected because it was the immune mediator measured in the most studies) in Fig. 7B.

All studies used clinician-collected CVLs. The CVL medium was saline in 19 studies, phosphate-buffered saline in 2 studies, and unspecified in another study. Thus, we did not have sufficient variation in methods to assess the effect of clinician- vs. self-collection or of lavage medium.

There was more variation in volume of CVL collected: 10 studies used 10 mL, 8 studies used 5 mL, 2 studies used 4 mL, 1 study used 2 mL, and 1 study did not specify. We were therefore able to compare concentrations of immune mediators recovered from 5 and 10 mL lavages (including all immune mediators that were measured in at least two studies at each volume). As shown in Fig. 7C, there was not a consistent difference between the concentrations of immune mediators detected in 5 and 10 mL CVLs (concentrations higher in 5 mL CVLs for 6/10 immune mediators, with p<0.05 for 1 of these; concentrations higher in 10 mL CVLs for the other 4 immune mediators with all p>0.05). This is illustrated at the level of individual studies in Fig. 7D, where the concentrations of CXCL8 detected in each study are shown stratified by CVL volume.

As an additional secondary outcome, we sought to determine whether one assay method yielded higher concentrations than the others. We compared the immune mediator concentrations detected by Luminex and MSD to ELISA (regardless of menstrual phase). For this analysis, we included all immune mediators that were measured using at least two assay methods, with each assay method being used in at least two studies.

As shown in Fig. 7E, Luminex gave lower total concentrations than ELISA for 12/15 immune mediators (p<0.05 for 3) and higher concentrations for 3/15 immune mediators (all p>0.05). MSD was mixed, with lower concentrations for 7/19 immune mediators (p<0.05 for 1) and higher concentrations for 12 (p<0.05 for 2 of these). This is illustrated at the level of individual studies in Fig. 7F, using CXCL8 as a representative example. As discussed in the Subgroup Analysis section above, the effect of menstrual cycle did not differ by assay method.

We next compared different methods of determining the menstrual cycle phase. Nine studies reported both days since the last menstrual period and serum progesterone levels. We used these studies to compare these two methods directly. Figure 8A shows all of the samples from those studies with their phases assigned by days since LMP (top) or by serum progesterone levels (bottom). Figure 8B shows that samples were rarely classified as opposite phases by the two methods: of the 535 samples that were assigned a phase (i.e., not undefined) by both methods, only 59 samples (11%) were assigned discordant phases. However, days since LMP lost many more samples to the undefined category. The two methods both designated 30 samples as undefined; an additional 130 were undefined by days since LMP, compared to only 62 by serum progesterone.

Menstrual phasing method did not have a consistent effect on the standard errors of the menstrual cycle effect sizes of individual immune factors across studies (Fig. S4A, difference between methods = 0.002, p = 0.87 by mixed model with study and immune factor as random effects, taken across all studies and immune factors). Within studies, the effect was consistent and dependent on sample size. In most studies, there were fewer undefined samples by serum progesterone than by days since LMP (for example, the studies Bradley, Cortez, and Hughes-unpublished). These studies tended to have lower standard errors in the analysis with phase determined by serum progesterone, consistent with the larger sample sizes in that analysis. Only one study had fewer undefined samples by days since LMP than by progesterone (Boily-Larouche). That study had lower standard errors in the analysis with phase determined by days since LMP. In addition, the effect sizes correlated well between the analyses performed with both phasing methods with Pearson r between 0.5 and 0.97 for all studies (not shown).

We next wished to determine whether immune mediator concentrations should be normalized to the total concentration of protein in the samples. Normalization to total protein did not have a consistent effect on the standard errors of the menstrual cycle effect sizes (Fig. S4B, difference between normalized and non-normalized = 0.011, p = 0.67, mixed model with study and immune factor as random effects, taken across all studies and immune factors). In most studies, the standard errors were very similar whether the analysis was performed on raw or normalized concentrations. In addition, the effect sizes were very strongly correlated between normalized and raw concentrations with Pearson r > 0.9 for all studies (not shown).