Chronic Migraine and Alzheimer’s Dementia: A Candidate Mechanistic Route, and the Evidence That Would Test It

1. Introduction

Dementia affected approximately 57 million people in 2019 and is projected to reach 153 million by 2050 [1]. The 2024 Lancet Commission identified 14 potentially modifiable risk factors accounting for around 45% of cases [2]. Migraine is not among them, and whether it belongs there has been raised repeatedly without being settled [3,4,5].

This review separates the two halves of that question. The mechanistic half asks whether a biologically coherent route runs from recurrent attacks to Alzheimer pathology. The epidemiological half asks whether the observational and genetic record can tell us if that route is traveled. The answers differ sharply, and saying so precisely — then specifying what would have to be measured instead — is more useful than another restatement of the association.

2. Methods

We searched PubMed, Google Scholar, and CrossRef (January 2020 – August 2026) using combinations of “migraine”, “chronic migraine”, “migraine with aura”, “Alzheimer’s disease”, “dementia”, “cortical spreading depression”, “glymphatic”, “brain insulin resistance”, “TNF-alpha”, “CGRP”, “brain age”, “Mendelian randomization”, and “chronic pain”, restricted to English-language studies in humans or translationally relevant animal models. Seminal work predating 2020 was added by hand-search. Following an internal audit in August 2026, every quantitative claim was re-derived from the source record rather than from an earlier draft; claims that could not be so derived were removed. Two preprints are flagged in the reference list.

Generative artificial intelligence (Claude, Anthropic) was used to assist with literature synthesis, with verification of references and reported values against the source records, and with language editing. No content was incorporated without verification against a primary source. All scientific claims, attributions, interpretations and conclusions were formulated and verified by the author, who takes full responsibility for the accuracy of the content. The tool is not, and could not be, an author.

3. What the Epidemiology Can and Cannot Support

An association is reported consistently: a 2025 meta-analysis of 11 cohort studies comprising 6 964 353 participants gave hazard ratios of 1.26 (95% CI 1.09–1.46) for all-cause dementia and 1.32 (1.26–1.38) for Alzheimer’s dementia [3]. What it cannot support is the more specific claim on which a cumulative-injury model depends.

3.1. Frequency Has Never Been Measured Against Dementia

The largest and most carefully constructed registry study — 44 195 patients risk-set matched to 44 195 controls on a time-dependent propensity score, mean baseline age 55.3 years, followed up to 16 years — reported hazard ratios of 1.30 (1.25–1.35) for all-cause dementia and 1.29 (1.23–1.35) for Alzheimer’s dementia, but defined exposure entirely from diagnostic codes, with no headache diaries, attack counts, or severity measures. Its authors close by stating that future studies need to explore whether characteristics such as severity or frequency affect the development of dementia [13].

One study measured frequency directly. In a clinical cohort with diary-confirmed diagnoses and interictal imaging, chronic migraine showed a brain-age gap of +4.16 years against −0.52 in controls (pairwise p = 0.010, Cohen’s d 0.49), with episodic migraine intermediate and not significant. Two qualifications belong with that figure: the three-group omnibus test did not reach significance (ANCOVA p = 0.053), and the model’s mean absolute error in controls was 6.26 years, larger than the gap itself [19].

The same study found no association between the gap and headache frequency (p = 0.89 in chronic migraine, 0.72 in episodic, 0.30 combined), migraine frequency, migraine duration, or chronic-migraine duration, and its authors state that a damage-accumulation model predicts exactly such a correlation [19]. Three things complicate a simple negative reading, all from the same paper. The mediation analysis was positive: segmented brain volume mediated the relationship between headache frequency and the gap across migraine groups combined. The authors note that frequency is already embedded in the chronic-versus-episodic categorization, truncating within-group variance; high-frequency episodic migraine was excluded by design, truncating it further. And they observe that migraine often ameliorates with age, so the largest structural effects may occur early and be invisible cross-sectionally. A direct null with a positive indirect path, in a cross-sectional dataset with a restricted exposure range, is not a refutation of dose-response; it is an argument that dose-response has not been measured in a design capable of detecting it.

3.2. Chronicity Has Never Been Stratified

The claim that chronic migraine carries risk beyond episodic migraine is untested. The meta-analysis most often cited for it performs no chronic-versus-episodic and no frequency stratification, and reports that other subgroup analyses could not be performed for lack of data [4]. A manuscript about chronic migraine therefore rests on an epidemiology that is almost entirely about migraine undifferentiated.

3.3. Aura Is Unresolved, Not Supportive

Depolarization is the established substrate of aura [22], so a model making it the initiating event predicts that aura carries the larger risk. The evidence shows neither that nor cleanly the reverse. The Mendelian-randomization analysis from which we take the thalamic mediation estimate reports an effect for migraine without aura (odds ratio 1.091, 95% CI 1.059–1.123, p = 6.95 × 10⁻⁹) and no corresponding effect for migraine with aura among its results [20]. The largest clinical cohort found the subtypes similar, but more than half its patients had no aura information recorded, and its authors call for re-investigation with accurate aura assessment [13]. Our position is that aura cannot serve as evidence for depolarization primacy anywhere in this paper — not because the evidence refutes it, but because no adequate aura-stratified evidence exists. If subclinical depolarization is common in migraine without aura, the clinical phenotype is in any case a poor proxy for depolarization burden. We adopt that consistently and do not invoke aura as support.

3.4. Direction Tracks Study Design

The most consequential pattern is methodological, and it can be quantified. The meta-analysis stratifying by design reports, for primary headache, an odds ratio of 1.49 (1.20–1.85) in case-control studies, 1.35 (1.25–1.45) in retrospective cohorts, and 0.82 (0.74–0.92) in prospective cohorts — significantly below the null. For migraine the same ordering holds: 1.59 (1.36–1.85), 1.37, and 0.88 (0.73–1.05), the last not significant [4].

Three features make the gradient hard to dismiss. Heterogeneity falls once studies are grouped by design — I² of 9.2%, 61.9% and 26.9% respectively, against 85.9% undivided — so the subgroups are internally consistent rather than pooling artefacts. Publication bias runs the same way: Egger’s test was significant among retrospective studies (p = 0.020) and not among case-control or prospective studies (p = 0.219, p = 0.522) [4]. The positive estimates come from the designs most exposed to ascertainment and reporting bias, the inverse estimate from the design least exposed to either. The authors’ own closing sentence is that most included prospective studies do not confirm the association, which might raise the suspicion of confounding [4].

A 2026 prospective cohort with adjudicated, imaging-supported outcomes is consistent with that subgroup rather than an outlier against it, reporting 0.70 (0.51–0.95) for all-cause dementia and 0.58 (0.40–0.85) for Alzheimer’s disease [26]; a 2018 review found no progressive cognitive decline at population level [34]. We previously called that 2026 cohort the central open problem. It is not: it is the newest point on a line already visible in a meta-analysis we cite. We also previously offered three defenses of it and withdraw two. Participants with migraine there had lower all-cause mortality, so competing risk biases their dementia incidence upward, not downward; and misclassification of the kind its instrument produces attenuates toward the null rather than generating a protective estimate. The defensible criticisms are narrower: exposure was lifetime rather than active migraine, probable migraine counted as non-migraine, and event counts did not permit stratification [26].

4. The Specificity Problem

If recurrent depolarization is the initiating and rate-limiting event, migraine should carry a dementia signal that pain conditions without depolarization do not. On the only meta-analysis stratifying by pain phenotype, it does not. Pooled hazard ratios were 4.47 (3.09–6.47) for trigeminal neuralgia, 2.77 (2.61–2.95) for fibromyalgia, 1.43 (1.06–1.92) for widespread pain, 1.31 (1.14–1.52) for osteoarthritis, 1.26 (1.16–1.36) for migraine, and 1.15 (1.09–1.21) for chronic non-cancerous pain, against a pooled chronic-pain estimate of 1.42 (1.23–1.64) [39]. Migraine ranks fifth of six; trigeminal neuralgia, involving no depolarization whatever, carries roughly three and a half times the risk. We treat this as the decisive objection and state plainly that the model cannot be defended on effect size. What we dispute is that the ranking measures what it appears to measure.

First, the source attributes its own top-ranked estimate to a different mechanism: its authors trace the trigeminal-neuralgia signal to depression, note that depression is an established dementia risk factor, and propose that trigeminal neuralgia raises risk by promoting it [39]. Second, the pooling will not support fine ranking — I² was 97.5%, Begg’s test significant (p = 0.036), and the search closed in September 2022, so none of the 2023–2026 evidence appears. Third, the geography is hard to read biologically: the pooled United States estimate was 1.12 (0.99–1.27) and the French 1.23 (0.88–1.73), neither significant, while the positive signal was carried by Asian claims databases [39]. A biological mechanism does not stop at national borders; one operating through diagnostic contact does. A second literature complicates the ranking from the other side: an increased brain-age gap has been reported in trigeminal neuralgia and osteoarthritis but not in chronic back pain [19].

Fourth, and most consequentially, exposure in every one of these phenotypes is a diagnostic code rather than a measurement of disease activity, and a code marks a moment of clinical contact rather than a period of disease. That moment sits at a different point in each condition’s natural history relative to the dementia observation window. Migraine peaks in the fourth and fifth decades and frequently remits; trigeminal neuralgia has a late onset and persists. A migraine code recorded at 60 will often mark a waning or former exposure, while a trigeminal-neuralgia code in the same person marks an active one. Any cumulative-injury signal is attenuated by that mismatch in migraine specifically, and the ranking inherits the attenuation.

This is not our conjecture. The Korean cohort’s investigators identified the problem in their own data and tested it, observing that a first migraine code in a cohort aged 40 to 79 might indicate an active period of long-standing migraine rather than new onset [13]. The association attenuated with time since diagnosis — 1.42 (1.34–1.51) at 0–5 years, 1.26 (1.14–1.40) at 6–10, 1.16 (1.05–1.28) beyond 10 — which they read as reverse causation or shared cause. But it remained significant beyond ten years, and remained significant when restricted to migraine first coded before age 60. Attenuation and persistence sit in the same dataset, and only the first is usually quoted.

One observation now unifies four difficulties. The design gradient, the cross-phenotype ranking, the attenuation with time since diagnosis, and the inverse prospective estimates are not four independent problems. In every case exposure is recorded as a diagnosis rather than measured as disease activity, and the recording sits at a different distance from the outcome window depending on design and phenotype. Retrospective claims data capture migraine during clinical contact, often while active; prospective cohorts ascertain lifetime history by questionnaire in late life, when migraine has frequently remitted. The missing design is neither: a prospective cohort ascertaining active mid-life migraine with attack counts, followed long enough for dementia to occur. To our knowledge it does not exist, and its absence rather than the presence of contrary findings is what leaves this question open.

The prediction that follows is the seventh test of this model. If the ranking reflects where the coded exposure sits relative to the outcome window rather than mechanism strength, conditioning on disease active within that window should compress it, raising migraine’s estimate relative to trigeminal neuralgia’s. If the ranking survives that conditioning, the mechanistic reading is correct and this model should be abandoned. The data sit in the same claims databases from which the current estimates derive. Two routes remain open meanwhile: either the shared driver is generic to chronic pain, in which case depolarization is not the root node and the model should be restated at the level of chronic nociceptive burden; or migraine’s contribution is mechanistically distinct despite not being numerically larger, which cannot be argued from effect size. We take the first as the default and treat the second as the hypothesis under test.

5. Relation to Prior Work

The components below are not all new. The neuroenergetic and inflammatory levels are substantially a compression of Del Moro and colleagues, whose 2024 synthesis already places tumor necrosis factor α at the pivot between the two diseases and closes by proposing shared prophylactic targets [5]. What it lacks is the glymphatic layer and depolarization as root node. The chain from depolarization through perivascular closure to neuroinflammation has been drawn elsewhere, and the glymphatic system in migraine has a standard review linking it to Alzheimer’s disease [42]. A 2026 systematic review sets the evidentiary ceiling plainly: across ten studies, perivascular-diffusion findings were heterogeneous, and no literature establishes these systems as causative even for migraine, let alone onward to dementia [41]. What we believe unoccupied is narrower: the single depolarization-rooted hierarchy spanning clearance, energetics and inflammation together; the argument that adjusting for sleep, mood and vascular comorbidity is over-adjustment for mediators; and the predictions as a set with refutation conditions attached.

6. The Candidate Mechanism

We propose that repeated depolarization initiates a three-level feed-forward network (Figure 1), offered as a hypothesis to be tested rather than inferred from the epidemiology above.

Level 1 — clearance and aggregation. Experimental depolarization produces near-complete closure of penetrating paravascular spaces, recovering over approximately 30 minutes; the tracer was a fluorescent dextran, amyloid was not measured, group sizes were small, and the animals anesthetized [15]. A 30-minute closure recurring on 15 or more headache days a month is the whole of the quantitative argument. Human evidence is thin and contested: a multimodal MRI study reported a greater brain-age gap in chronic migraine using a coupling measure its authors call an indirect proxy [23], while a perivascular-diffusion study found the index raised rather than lowered, in the right hemisphere only and without surviving correction for multiple comparisons [33]. Separately, a multiscale model calibrated on thioflavin-T assays predicts that depolarization-associated contraction of the extracellular space accelerates amyloid-β 42 aggregation, with recurrent events amplifying oligomer accumulation — a computational result without animal or human validation, whose interest is that it makes attack frequency rather than severity the governing variable [35].

Level 2 — neuroenergetic deficit. Downregulated neuronal and astrocytic insulin receptors, impaired lactate shuttling and reduced glycogen synthesis produce an interictal energy gap exposed whenever cerebral demand rises [5,16]; metabolic-first accounts of Alzheimer’s disease converge on the same substrate [9,10]. The deficit is unlikely to be sex-neutral, given perimenopausal effects on cerebral glucose metabolism and mitochondrial bioenergetics [27,28]. Our contribution here is ordering rather than content, and the ordering is hypothesized.

Level 3 — transduction. This is the best-supported link. Depolarization opens neuronal pannexin-1 megachannels, releasing ATP and HMGB1; pannexin-1 opening activates the neuronal NLRP3 inflammasome, and HMGB1 engages toll-like receptors 2 and 4 on astrocytes and microglia. Inflammasome activation is exclusively neuronal, detectable within 15 minutes, and blockade at each node attenuates trigeminovascular activation [31,32]. The same signal is independently implicated at the Alzheimer end, driving cascades that promote tau hyperphosphorylation and amyloid accumulation [38]. Two literatures describe one mechanism from opposite ends; human confirmation is the open flank. We previously presented genetic evidence as establishing tumor necrosis factor α as the specific mediator. That does not survive contact with the source: migraine was modeled against 35 risk factors in separate adjustments, and the association survived neither adjustment for tumor necrosis factor α (1.2411, 0.8352–1.8443, p = 0.2852) nor for diastolic blood pressure (1.4120, 0.8487–2.3493, p = 0.1840), which the source’s abstract foregrounds equally [7]. No mediation analysis was performed and no mediation fraction is reported.

The levels are proposed to reinforce each other. One tension must be owned: if the loops are strong, removing any single level should be insufficient, yet predictions 4 and 5 test exactly single-level removal. Either the loops are weaker than proposed, or depolarization retains a rate-limiting privilege the other levels lack. We take the second position, and it is a commitment rather than an observation.

7. Comorbidities and the Over-Adjustment Problem

Prevailing practice adjusts for sleep disorders, depression and cardiovascular disease as independent confounders. Each has an independent etiology and each can trigger or chronify migraine [4,12]. But once established they share mechanistic surfaces with levels 1 to 3 (Table 1). If they lie on the causal path, adjusting for them is over-adjustment for mediators and the adjusted estimate is not the quantity of interest. This cuts both ways: published adjusted estimates may understate a true effect, and we cannot claim an independent effect of migraine while arguing that the comorbidities cannot be separated from it. What follows is not an independence claim but a call for mediation analysis with prespecified causal structure, which the literature does not provide.

8. Predictions, with Refutation Conditions

• Clearance. Direct kinetic measurement should show impaired clearance in chronic migraine, scaling with monthly headache days. Refuted if a study of at least 60 patients shows no monotonic relationship, or one in the opposite direction, with a confidence interval excluding a 10% difference between the highest and lowest frequency tertiles. Existing indirect evidence is null, opposite and non-committal [19,23,33]; frequency must be sampled across its full range rather than dichotomized.
• Thalamic trajectory. Longitudinal thalamic atrophy should exceed matched controls and predict cognitive decline. Refuted if a prospective cohort of at least 200 patients followed five years shows a difference whose confidence interval excludes 0.1% per year. The mediation estimate we cite derives from an analysis reporting annual thalamic atrophy of −9.507 cm³, which is not biologically possible given a thalamic volume of roughly 7 to 8 cm³ per hemisphere; the unit appears misstated and the fraction should be treated as provisional [20].
• Inflammatory specificity. Formal mediation analysis with pleiotropy-robust estimators should show tumor necrosis factor α carrying a mediated fraction that blood pressure and body-mass index do not. Refuted if either carries an equal or larger fraction — on the available adjustment pattern, the more likely outcome [7].
• Prophylaxis, legacy agents. Sustained midlife prophylaxis should attenuate later Alzheimer’s incidence at a 10 to 20 year lag. This needs a protocol rather than an aspiration: an active-comparator new-user design with one prophylactic class against another, a negative-control outcome, and a prespecified effect size. Refuted if a hazard ratio confidence interval excludes 0.95 in a cohort with adequate exposure duration.
• Prophylaxis, CGRP-pathway agents. Because these entered large-scale use only from 2018, no dataset supplies the follow-up required; first-decade results should not report before approximately 2038. Claims of neuroprotection resting on one- to two-year follow-up must not be conflated with disease modification.
• Subtype. If clinical aura marks depolarization burden, aura should carry the larger effect; it does not [20]. We therefore treat clinical aura as a poor proxy, and the prediction concerns depolarization measured directly: refuted if a cohort with electrophysiological or imaging quantification shows no relationship to subsequent amyloid or tau accumulation. Aura is also associated with patent foramen ovale [37], so subtype analyses should stratify for shunt status.

9. The Cheapest Decisive Test

No published work reports Alzheimer-type biomarkers in chronic migraine. A cross-sectional comparison of plasma phosphorylated-tau 217 and the amyloid-β 42/40 ratio between people aged 45 to 65 with chronic migraine, with episodic migraine, and without headache, with cumulative attack burden modeled continuously and sleep measured by actigraphy rather than assumed, would cost a fraction of any imaging program. The episodic group is essential: it controls for the diagnosis and isolates attack burden. If a difference of the implied size is excluded, the hypothesis is substantially weakened; if a gradient appears, the frequency question the entire epidemiology has failed to address is finally on the table. Blood biomarkers are validated for cognitively impaired populations in specialist care, not for midlife screening; such a study would measure group differences, not diagnose individuals.

10. The CGRP Question

Two hypotheses compete and the evidence is balanced. On the first, calcitonin gene-related peptide is an endogenous protector whose long-term blockade could leave the aging brain vascularly under-defended: in a rat model of chronic cerebral hypoperfusion, capillary constriction preceded and drove amyloid accumulation, and supplementation prevented that constriction and improved cognition [40]. Post-marketing pharmacovigilance has identified cerebrovascular reporting signals for individual monoclonal antibodies but not gepants [29]; these are disproportionality measures, not incidence. On the second, the peptide drives the proximal inflammatory cascade and blockade aligns with prevention [21,30]. The two predict divergent cognitive trajectories under chronic blockade, testable in linked-claims data as ten-year windows open from 2028. The current evidence favors neither.

11. Limitations

The principal limitation is stated in the body rather than confined here: the best-controlled designs do not support the association on which this model depends, and three features we previously offered as support are untested, untested and unresolved. Beyond that: Mendelian randomization gives direction-of-effect inference, not proof, and horizontal pleiotropy is particularly plausible because migraine-associated variants are enriched in vascular and ion-channel genes that plausibly act on dementia risk independently of attacks. Level 1 rests on rodent work with small groups and a computational model without in-vivo validation; level 2 on imaging and peripheral surrogates; level 3 has strong preclinical support and no human confirmation. The relationship between attacks and depolarization events is assumed rather than demonstrated, and the dose-response argument depends on it. This is a single-author narrative review without primary data; its contribution is the specification of a test, not the result of one.

12. Conclusions

The proposition that chronic migraine contributes causally to Alzheimer’s dementia is not supported by the best-designed evidence, and the reason is not that the question has been answered against it. Attack frequency has never been related to dementia incidence in any cohort. Chronicity has never been stratified. No biomarker study exists. The one study that measured frequency directly found a null direct association and a positive indirect path in a design with a truncated exposure range. Meanwhile the reported association falls monotonically as designs improve and inverts in prospective cohorts — which is what one would expect either if the association is an artefact of diagnostic contact, or if the exposure that matters has been mismeasured in every study so far. Those readings make different predictions and can be separated. A mechanistically coherent route exists and deserves testing, but it should be described as untested.