Patent Foramen Ovale, Ischemic Stroke and Migraine: Systematic Review and Stratified Meta-Analysis of Association Studies

Modifiable risk factors for ischemic stroke may be targets for primary or secondary prevention, especially if the strength of their association suggests that they might be causal. Up to one third of ischemic strokes are ‘cryptogenic’ because a recognized cause is not identified [1]. One putative cause of cryptogenic ischemic stroke is paradoxical embolism through patent foramen ovale (PFO) [2], itself a consequence of failure of the septa on either side of the fetal interatrial shunt to fuse after birth. Moreover, suggested associations between PFO and migraine are of interest because migraine has also been considered a risk factor for ischemic stroke [3].

Non-blinded and non-randomized interventional studies of percutaneous PFO closure have suggested benefits of PFO closure on recurrent cryptogenic ischemic stroke [4,5,6] and refractory migraine [7,8,9,10], but randomized controlled trials have not confirmed these effects [11,12]. This suggests the possibility that the reported observational data, like the uncontrolled trials, may have been biased or confounded. While the randomized trials may have been underpowered to detect a benefit of closure, this may have been driven by the original estimated effect size being based on a stronger association than is actually the case.

In this context, it is timely to review the observational studies and the way in which previous meta-analyses addressed methodological biases when pooling estimates of association. To date, systematic reviews have suggested significant associations between migraine and ischemic stroke [odds ratio (OR) 2.2; 95% confidence intervals (CI) 1.6–3.0] [13], PFO and ischemic stroke (OR 2.9; 95% CI 2.1–4.0) [14], and PFO and migraine (OR 2.5; 95% CI 2.0–3.1) [15]. However, the degree of bias in observational data differs between studies, and not accounting for these differences may lead to propagation of these biases in the pooled estimates [16]. One approach to this issue is to stratify analyses based on methodological design, thereby reducing heterogeneity in observational studies in the meta-analysis.

Here, we set out to refine the interpretation of the current literature in order to (1) quantify the interrelated associations between migraine, PFO and stroke, and (2) to inform the design of randomized trials of PFO closure. Therefore, we undertook a series of systematic reviews and meta-analyses investigating the pairwise associations between a past history of migraine, PFO and the occurrence of cryptogenic ischemic stroke or all ischemic stroke (fig. 1), applying consistent eligibility criteria across all three relationships. We specifically gave priority to population-based studies, prospective cohort designs and studies that had matched or adjusted for known confounders for stroke, with a view to stratifying by study design and restricting meta-analyses to methodologically similar population sampling frames.

The protocol followed the Meta-analysis of Observational Studies in Epidemiology (MOOSE) guidelines [17] (online suppl. checklist; for all online supplementary material, see www.karger.com?doi=10.1159/000341924; study characteristics of all included studies are given in full, along with referenced PRISMA flowchart detailing excluded studies).

We searched for cohort, case-control or cross-sectional studies unrestricted by demographic features or language of publication, which had been conducted in population or hospital settings, and which compared the associations between any two relationships (fig. 1). We required studies defining ischemic stroke by WHO criteria (supported by radiological findings) and cryptogenic ischemic strokes defined as non-lacunar clinical syndromes without carotid stenosis >50%, cardiac arrhythmias or structural heart disease (excluding PFO). To be included, we required studies that did not report an outcome of cryptogenic ischemic stroke but instead report total first-ever or recurrent ischemic stroke with adjustment for known risk factors for ischemic stroke (specifically hypertension, diabetes, hypercholesterolemia, atrial fibrillation, smoking, alcohol, family history). We included studies defining migraine according to the International Headache Society (IHS) classification [18,19]. PFO must have been diagnosed by contrast transthoracic echocardiogram (TTE), transesophageal echocardiogram (TEE) or transcranial Doppler (TCD) ultrasound during Valsalva maneuver. Cerebral infarction in the context of complex migraine is a well-recognized clinical phenomenon, and so studies reporting migrainous infarction were not considered further. We did not include data from placebo arms of randomized trials because interventional studies are not required to use a consistent sampling frame. We excluded studies of children, pregnant women and those reporting associations within family pedigrees.

D.D. searched Medline from 1950 on PubMed, Embase from 1980 on NHS Evidence Health Information Resources, and Science Citation Index on Web of Science from 1950, all to April 2012. We sought journal publications as well as conference abstracts and contacted authors of abstracts to determine whether further data had been published. In addition, the bibliographies of included articles and other systematic reviews were screened.

D.D. used comprehensive textword and Medical Subject Heading (MeSH) to find relevant studies in humans (full search terms given in MOOSE checklist, online suppl. material). Four reviewers (D.D., J.G., P.W., B.S.) independently screened articles by title and abstract after de-duplication through first author surname, title and first page number with those for full-text review, and subsequent inclusion agreed by consensus. We sought every publication relating to each included study using a search strategy based on the full name of the study or its acronym, as well as forward searching using citation index databases.

Two reviewers independently extracted study data using standardized forms developed after piloting and critically appraised each study using the Newcastle-Ottawa scale (NOS) [20]. The NOS assesses population selection, comparability of cases and controls, and adequacy of outcome assessment (including outcome ascertainment and attrition). In keeping with the MOOSE guidelines, we did not use the NOS to determine study inclusion but to guide the classification of studies by similarities in design.

We classified a hierarchy of population sampling methods from lowest to highest risk of bias (online suppl. table S1), where ‘population-based’ refers to a sampling frame including all subgroups within a geographical population, unrestricted by demographic or clinical features, ‘community-convenience’ indicates selection from outpatient participants, and ‘hospital-convenience’ denotes participants recruited only from hospital inpatients. We considered cryptogenic ischemic stroke and ischemic stroke as separate outcomes.

J.G. and P.W. conducted the statistical analyses using STATA version 10.1. Where multiple estimates of association were reported, the most adjusted estimate was used in analyses. Summary statistics for hazard ratios (HR), relative risks (RR) or OR and their accompanying 95% CI were calculated using a Mantel-Haenszel fixed-effect model. Statistical heterogeneity was assessed with the I² statistic.

We identified 7,686 studies (online suppl. MOOSE checklist). We reviewed the full text of 116 studies, and excluded studies are fully referenced in the PRISMA flow diagrams (online suppl. fig. S1–S3). Supplementary table S2 gives further details of any study considered in any previous systematic reviews and meta-analyses, if any such study was excluded here. Finally, we included 37 studies in this review. We have separately summarized study characteristics, their results (and their meta-analysis, where appropriate) and risk of bias according to the association being investigated. Full details of all reviewed studies are given in online supplementary tables S3–S5. Table 1 shows the characteristics of only those included studies with either population-based sampling frames or a prospective design.

We identified one prospective and one population-based case-control study of women with migraine with aura with conflicting results (fig. 2).

Prospective Studies. The Women’s Health Study (WHS), which included 27,840 female health care professionals aged >45 years, was the only prospective study in this section to meet all inclusion criteria [21]. The most fully adjusted model estimated an association with ischemic stroke (HR 1.2; 95% CI 0.9–1.7 for migraine and HR 1.9; 95% CI 1.2–3.1 for active migraine with aura).

Population-Based Case-Control Studies. One case-control study used a population-based sampling frame [22]. In women aged 15–49 years, 192 participants with migraine with aura were not associated with cryptogenic stroke (adjusted OR 1.4; 95% CI 0.9–2.0). In a small subgroup of these women with known PFO, there was no evidence of a different association (adjusted OR 2.1; 95% CI 0.8–5.3; n = 21) compared to the group known to not have PFO (adjusted OR 1.5; 95% CI 1.0–2.2; n = 142).

Non-Population-Based Case-Control Studies. Figure 4a shows the estimated associations for any history of migraine and ischemic stroke, stratified by study design. The reported association of one study comparing cryptogenic stroke cases with community-convenience controls (OR 1.1; 95% CI 0.5–2.4) [23] was somewhat weaker than the association in a study comparing ischemic stroke cases [24] with community-convenience controls (OR 1.7; 95% CI 1.1–2.8) than the pooled risk estimate of studies of ischemic stroke with hospital-convenience controls [25,26,27,28] (pooled OR 2.0; 95% CI 1.4–2.8; I² = 58%). Figure 2b shows the meta-analysis of studies reporting migraine with aura subtypes [25,26,27,29]. This yielded a pooled OR of 2.9 and 95% CI of 1.7–4.9 (I² = 35%). Failure to account for atrial fibrillation was a feature common to all these studies.

Pooled results from prospective and population-based studies did not demonstrate an association between PFO and ischemic stroke (fig. 3).

Population-Based Prospective Studies. We identified 2 population-based cohort studies in essentially stroke-free populations where ischemic stroke adjusted for known vascular risk factors was the main outcome (fig. 3) [30,31]. The pooled HR for risk of primary ischemic stroke in relation to PFO was 1.6 with 95% CI of 1.0–2.5 (I² = 0%). The other prospective studies reported no statistically significant association (where p < 0.05), either individually or if pooled (OR 1.0; 95% CI 0.6–1.8; I² = 0%) (fig. 3).

Population-Based Case-Control Studies. Three studies compared cryptogenic ischemic stroke with healthy controls sampled from the general population, yielding a pooled OR of 1.3 and 95% CI of 0.9–1.8 (I² = 0%) (fig. 3) [32,33,34]. Two studies performed subgroup analyses to assess the association of larger shunts with stroke [32,33]. Pooling the association across studies yielded an OR of 1.3 and 95% CI of 0.8–2.1 (I² = 79%).

Non-Population-Based Case-Control and Cross-Sectional Studies. A total of 10 studies compared cryptogenic ischemic stroke with: (1) healthy community-convenience controls [35,36]; (2) controls undergoing echocardiography for non-stroke indications [37,38,39], or (3) ischemic stroke of known cause, adjusted for other vascular risk factors [40,41,42,43,44]. When subdivided by a population sampling frame, the pooled ORs for community-convenience and hospital-convenience controls were 1.9 (95% CI 1.2–3.1; I² = 39%) and 2.8 (95% CI 2.1–3.6; I² = 27%), respectively (fig. 3).

We found one population-based study, and this did not show an association between PFO and migraine (fig. 4).

Cross-Sectional and Case-Control Studies. In addition to the stroke outcomes above, the population-based NOMAS study [45] reported the association between PFO and migraine as OR of 1.0 (95% CI 0.6–1.6), and this was not different when only considering migraine with aura (OR 1.0; 95% CI 0.7–1.7). Garg et al. [46] reported no association with PFO (OR 1.0; 95% CI 0.6–1.7) in 144 stroke-free migraineurs recruited from a tertiary clinic who were compared to age- and sex-matched controls selected from a register of healthy volunteers. The other case-control study used convenience controls, and the estimate for any history of migraine and PFO was OR of 3.9 (95% CI 1.4–11) (fig. 4) [47]. Two further abstracts were identified that reported population-based associations, and full publication of both studies is awaited [48,49].

For each pairwise association between migraine, PFO and ischemic stroke, there is inconsistent evidence, and the strength of the reported associations is dependent on study design. Overall, associations apparent in case-control studies were not evident in the more generalizable population-based studies with sufficiently valid assessments of exposure and outcome.

Migraine and Ischemic Stroke. In the subset of younger women with migraine with aura and cryptogenic stroke, there is no clear association [22], and there are no studies with a strict population-based sampling frame describing this relationship in men (of any age). However, there is some evidence from a prospective study suggesting an association in female health workers aged >45 years with migraine with aura and ischemic stroke [21], though atrial fibrillation was not accounted for despite a separate report demonstrating an association between this and stroke [50].

PFO and Ischemic Stroke. Cohort studies did not demonstrate an association between PFO and first-ever [30,31] or recurrent ischemic stroke risk [51,52], findings which are supported by three case-control studies comparing PFO in cryptogenic ischemic stroke with population-based controls [32,33,34].

PFO and Migraine. The only population-based study investigating the association between PFO and migraine with or without aura was cross-sectional and found no association [45].

This is the first review on this topic to stratify analyses on methodological grounds, which may explain why our conclusions differ from earlier meta-analyses. In applying a consistent approach to article selection and critical appraisal, we have been able to assess the evidence systematically across the three relationships. Our approach to meta-analysis has had the advantage of yielding low statistical measures of heterogeneity (I²). While there are other procedures for addressing heterogeneity [53], our method has sound face validity. By identifying the observational studies at least risk of bias, we base our conclusions on the most epidemiologically applicable data.

For studies of migraine and ischemic stroke, there was inconsistency in controlling for possible confounders, in particular atrial fibrillation. Accordingly, the one prospective study to report a positive association (the WHS) should be considered with caution in this light.

The two population-based prospective studies of PFO and first-ever stroke risk had very similar results and did not demonstrate an association. It is possible that longer follow-up might have detected a small contribution of PFO to ischemic stroke risk. However, given the age-related incidence of more well-recognized risk factors for ischemic stroke, the power to detect an attributable risk from PFO probably diminishes with time. The population-based case-control studies were in agreement, though the findings from Roijer et al. [34] are problematic because the age matching was unsuccessful, with controls being younger. In addition, although ‘major potential cardio-embolic sources’ were excluded, a higher frequency of atrial fibrillation was reported in the cases. Despite these possible biases, the overall conclusions are acceptable in the context of the other population-based studies.

The only population-based data on the association between migraine and PFO assessed migraine in those reporting a past history of headache during a screening interview. Given the mean age of the participants (69 years), it has been suggested that recall bias may have caused an underestimation of migraine history, though the reported prevalence of migraine is consistent with other estimates from the general population [45]. Finally, because the pooled results from population-based studies did not demonstrate any clear associations, it was not possible to examine for any potential interactions or effect modifiers.

There are three main reasons that population-based observational studies may consistently report no association between migraine, PFO and cryptogenic stroke. Firstly, any misclassification of exposure status may bias any true association towards the null. In particular, studies employing TTE may have not detected PFO as reliably as more sensitive methods. Secondly, population-based studies may themselves been underpowered to detect the population-attributable fraction for stroke from PFO. Thirdly, the null hypothesis that PFO is not associated with stroke or migraine in unselected populations cannot be rejected. Therefore, there are general implications for trials considering PFO closure in terms of how to assess PFO, which subpopulations to sample and the sample size required to demonstrate a modifiable effect.

Previous meta-analyses have separately examined the relationships reviewed here. With regard to migraine and ischemic stroke, two recent systematic reviews have been conducted [13,54]. The eligibility criteria in these reviews did not specifically focus on cryptogenic stroke and included a larger number of studies. One review pooled – with equal weighting – evidence across cohort (HR) and case-control studies (OR), finding an association measure of 2.2 (95% CI 1.5–3.0) [13]. The other meta-analysis adopted similar methods but included a separate analysis restricted to studies at low risk of bias as determined by the authors (fewer than three ‘poor’ scores in different areas of the study design), in essence assigning an overall score for quality [54]. However, even in the analysis restricted to high-quality studies, the statistical heterogeneity was still sufficiently high to require a random-effects meta-analysis (I² = 68%). Because this in effect combines studies across different populations, it is no longer clear to which population the pooled estimate then applies. Assigning an overall quality score does not in itself address the biases within individual studies. We avoided this by stratifying our analyses where risk of bias due to study design and population sampling is likely to be similar across studies.

Two systematic reviews on the relationship between PFO and cryptogenic ischemic stroke have been published [14,55]. Both of these only examined case-control studies, thus do not account for evidence derived from prospective studies, and their results are consequently less conservative than ours, where the PFO-stroke association was estimated as OR of 3.0 (95% CI 2.0–4.3) for studies of PFO in cryptogenic stroke compared to non-stroke controls [55]. For the PFO-migraine association, another meta-analysis demonstrated an OR of 2.5 (95% CI 2.0–3.1; PFO in migraineurs) and an OR of 5.1 (95% CI 4.7–5.6; migraine in PFO patients) [15], though the systematic review predated the NOMAS report and also acknowledged that publication bias was likely. Two abstracts await full publication but appear to report associations in population-based samples. The Genetic Epidemiology of Migraine (GEM) study found migraine and PFO to be related only where PFO was also associated with an atrial septal aneurysm [48]. CAMERA is a MRI substudy of GEM, reporting a significant association in migraine with aura and PFO (OR 2.1; 95% CI 1.1–3.9) [49]. However, it is unlikely that pooling the preliminary result with the NOMAS data would substantially alter these conclusions given that NOMAS is a much larger study.

The findings of this systematic review have implications for trials of PFO closure. If the underlying association is weaker than previously recognized (or non-existent), detecting a treatment effect may require enormous sample sizes (or may be impossible). However, if randomized trials demonstrate a treatment effect, this would provide evidence not only for an association between PFO and ischemic stroke or migraine, but one that is causal.

Our findings may account for the results of the CLOSURE trial [12]. The trial had an event rate (TIA and stroke) of 6% in the control arm and 2% in the closure arm. Also considering adverse effects of treatment, there were 7.7 and 5.9% events in the control and intervention groups, respectively (p = 0.30). Regardless of the concerns surrounding peri-procedural atrial fibrillation in this and other cohorts [56], one interpretation of the trial is that the underlying association is too small to be modified by an intervention with this efficacy.

The Migraine Intervention With STARFlex Technology (MIST) trial is the only double-blinded sham-controlled intervention trial in patients with refractory migraine with aura [11]. Though questions remain in respect of the possibility that interatrial shunts were not completely closed [57,58], the reported data suggest no benefit of PFO closure on migraine. It has been proposed that MIST participants were not representative of those in whom treatment benefit was observed in uncontrolled trials (i.e. those migraineurs spontaneously improving after PFO closure for cryptogenic stroke) [59]. This observation is supported by the community-based studies finding no association between migraine and PFO in stroke-free populations identified in this review.

Our findings do not support an association between PFO in patients with cryptogenic ischemic stroke. Because the weak association between PFO and ischemic stroke in methodologically rigorous studies questions the causal role of PFO, randomized controlled trials of PFO closure constitute an essential test of the causal relationship. However, the weakness of the association in observational epidemiological studies, as well as the effectiveness of antithrombotic drugs in the secondary prevention of ischemic stroke [60], means that ongoing and future trials of PFO closure will need to be very large. PFO closure may be more effective in subgroups with a stronger association with ischemic stroke (and therefore at highest risk of future cardioembolism), possibly in persons with very large PFO shunts [32] or PFO associated with atrial septal aneurysm [48].

D.D. is funded by a Wellcome research training fellowship. J.G. is sponsored by a joint MRC/GlaxoSmithKline industrial CASE studentship. P.W. is supported by a non-clinical PhD studentship from the British Heart Foundation. B.S. is funded by the Joint European Post-Doctoral Programme: The European Research Area in Ageing (ERA-AGE) Network FLARE Programme. R.A.-S.S. was funded by a UK Medical Research Council clinician scientist fellowship.

Study characteristics of all included studies are given in full, along with referenced PRISMA flowchart detailing excluded studies. The MOOSE checklist is also submitted.

The authors have no conflict of interest to declare.