A Stroke of Insight? Possible Bidirectional Relationship Between Stroke and Epilepsy

Abstract

Background: Observational studies have established a relationship between stroke and epilepsy. While most studies have reported an increased risk of epilepsy following a stroke, fewer have concluded that a diagnosis of epilepsy increases the risk of future strokes. Precisely describing the causal relationship between epilepsy and stroke has clinical and public health implications. Methods: We performed a two-sample Mendelian Randomization (MR) to analyze the relationship between epilepsy and stroke. We identified genetic instruments from the Stroke Multiancestry Genome-Wide Association Study [GCST005838 (67,162 cases and 454,450 controls)] and the International League Against Epilepsy Consortium on Complex Epilepsies [Generalized epilepsy (GE): GCST007343 (n_case=3769, n_control=29677), Focal epilepsy (FE): GCST007352 (n_case=9671, n_control=29677)]. We used a significance threshold of p-value <5 × 10⁽⁻⁵⁾ for genetic instrument identification. We included the following epilepsy phenotypes: GE, FE, childhood absence epilepsy (CAE), juvenile absence epilepsy (JAE), juvenile myoclonic epilepsy, GE with tonic-clonic seizures, FE with hippocampal sclerosis (focal HS), and focal lesion-negative epilepsy. Linkage disequilibrium reference panel of 1000 Genome Project pruning with Pearson correlation r² < 0.2 was applied to ensure that the analysis is restricted to independent variants. We used Steiger filtering for reverse causality and Z-scores for GE and FE. We employed the following methods for causal effect estimation: Inverse-variance weighted (IVW), MR-Egger, MR-RAPS, and MRPRESSO. Results: For the association between GE and stroke (GE > stroke), stroke and GE (stroke > GE), FE and stroke (FE > stroke), and stroke and FE (stroke > FE), the number of single nucleotide polymorphisms was respectively 253, 213, 187, and 210. GE was associated with an increased risk of stroke [IVW, 95% confidence interval: 0.058 (0.028-0.088)], and stroke was also associated with an increased risk of GE [IVW: 0.059 (0.026-0.092)]. FE was associated with an increased risk of stroke [IVW: 0.066 (0.029-0.102)] and stroke was associated with an increased risk of FE [IVW: 0.037(0.005-0.068)]. The analysis by epilepsy subtypes revealed an increased risk only among patients with stroke and for the following epilepsy subtypes: JAE, CAE, Focal HS, and Focal lesion negative epilepsy. Conclusion: Our findings support a bidirectional relationship between stroke and epilepsy. Studies on post-stroke epilepsy should account for this bidirectional relationship.

Commentary

Understanding the complex, potentially bidirectional relationship between epilepsy and its etiologies involves many challenges. For example, stroke represents a major cause of epilepsy, especially in older adults with vascular risk factors. Still, stroke survivors may have other seizure risk factors that could influence epilepsy incidence. Similarly, whether epilepsy itself increases the risk of stroke due to the underlying abnormal electrical network, independent of traditional risk factors (eg, enzyme-inducing antiseizure medications, smoking, hypertension, obesity, physical inactivity), is less well-accepted.

Chen et al ¹ recently examined the above bidirectional possibilities. They asked: (1) Does stroke cause epilepsy? And (2) Does epilepsy cause stroke? They used two datasets: (1) The Stroke Multi-Ancestry Genome-Wide Association Study. This includes about 520,000 individuals (67,000 with epilepsy and 454,000 without epilepsy), with a mixture of origins from Europe, Asia, Africa, and Latin America. And (2) The International League Against Epilepsy Consortium on Complex Epilepsies. This includes about 44,000 individuals (15,000 with epilepsy and 29,000 without epilepsy), almost all European.

Their analyses employed a technique called Mendelian Randomization (MR), which considers genetic variation as a type of Instrumental Variable. ² The idea is that many unobserved confounders likely distort the relationship between the exposure (X) and outcome (Y) of interest. One approach would be to measure and analytically incorporate as many confounders as possible. Though, datasets often do not capture all relevant confounders, nor are all relevant confounders known. An alternative approach is to identify Instrumental Variables (in the case of MR, random genetic variants) theorized to exogenously manipulate X (the “relevance” assumption) but itself have no other influence on Y (the “exclusion restriction” assumption), thus breaking the link between X and Y confounders. Additional assumptions include “exchangeability” (there are no unmeasured confounders between the Instrumental Variable and Y) and one additional identifying assumption such as “monotonicity” (the effect of the genetic variant on X is in the same direction for everyone). One could then interpret analyses as the variation in Y that is due to random genetic-induced variation in X, independent of X–Y confounders (even when one has not measured or incorporated X–Y confounders). They chose genetic variants based upon significance in Genome Wide Association Studies—those (∼200) Single Nucleotide Polymorphisms (SNP) that were correlated with either stroke or epilepsy with p < .00005.

Their main result was that both generalized epilepsy and focal epilepsy predicted stroke. Also, stroke predicted both generalized epilepsy and focal epilepsy. In their main inverse-variance weighted results (a way to pool results across all ∼200 SNPs to estimate a single X–Y relationship), log-odds ranged from about 0.037 to 0.066, which after exponentiating translates into odds ratios of approximately 1.038–1.068.

The authors conclude that the data support a bidirectional causal relationship between stroke and epilepsy. Such a conclusion could have numerous implications. For example, while it was already suspected that stroke increases the risk for epilepsy, underscoring this relationship could make clinicians watch these patients more closely, within the context of known risk factors for post-stroke epilepsy. ³ Likewise, if epilepsy leads to stroke, this could clue clinicians into the need for closer monitoring, a lower threshold for primary cardiovascular prevention efforts, and increased counseling about reasons to dial 911 for stroke symptoms.

However, assumptions for an MR analysis are unclear here, thus it seems premature to label estimates as causal. A valid Instrumental Variable should strongly influence X, and they did choose SNPs with very low p-values. But GWAS studies search through an enormous number of candidate SNPs thus ensuring a smaller number of spuriously significant findings that would be unlikely to be reproduced in external samples. The analyses do include sensitivity methods that either adjust for outlier variants or relax pleiotropy assumptions. But the core assumptions characterizing a valid Instrumental Variable analysis are either dubious or not explicitly clear. There is no biological rationale for why any of the specific chosen SNPs would influence X beyond noise (“relevance” assumption). Rather, the Figures support Instruments based on noisy significance (“winner's curse” exaggerating effects) rather than biological theory or content knowledge, judging from bimodal distributions skipping over the null region. There is also no biological argument that any of those same genetic variants should not be expected to have other paths by which they influence Y (“exclusion restriction” assumption), or other understanding for why variants might have the same direction of effect across populations (“monotonicity” assumption).

Magnitudes are also small. For example, say 1% of people without a stroke have epilepsy. To say that stroke confers a ∼4% relatively increased odds of epilepsy would be to say that stroke increases the risk of epilepsy from 1.00% to about 1.04%, which is both far below any possibly clinically relevant threshold and well below any other estimates in literature. And without a clear time-course about whether seizures or stroke came first, with selection bias from lifetime genetic exposure among survivors, and without description as to what duration of follow-up does the odds ratio correspond, results are further difficult to interpret.

The results also at times contradict clinical intuition. For example, while it is already counterintuitive that stroke would cause generalized epilepsy, it is further difficult to accept results that stroke would have the same or even less of an effect on focal epilepsy as on generalized epilepsy. The above concerns raise the likelihood that effects are driven by synthesizing weak, noisy instruments chosen by p-values rather than biological plausibility, each with small but consistent directions of bias, with contribution from differential measurement error (patients with epilepsy are more likely to get an MRI thus detecting occult/incidental lesions; patients with one disease are more likely to seek care for other diseases), leading to underestimating true effects and overestimating false effects. A negative control, ⁴ more biological theory about the specific SNPs, or else using separate datasets to choose SNPs then estimate outcomes could all help future studies.

Studying factors contributing to development of stroke, epilepsy, and their related pathways are critical to enhancing our understanding of which patients to target for individualized counseling and prevention efforts. Chen et al have completed an interesting example illustrating modern techniques at our disposal to help understand the causes and effects of epilepsy. These results could support the known role of stroke in epileptogenesis and that patients with epilepsy may have many risk factors that lead to stroke to direct counseling. Future studies can more carefully seek to disentangle the various pathways at play, such as the degree to which such pathways are due to mediating indirect effects and by what variables. Though careful attention to assumptions and articulating well-defined counterfactuals will be critical to robust inferences. ⁵