Premonitory symptoms in migraine: a systematic review and meta-analysis of observational studies reporting prevalence or relative frequency

The definition of premonitory symptoms (or prodrome) used above is provided by the International Classification of Headache Disorders (ICHD) (Supplementary Table 1) [3, 8‐10]. In the four iterations of ICHD so far published, the term premonitory symptoms is recommended in the first three (ICHD-1, ICHD-2, ICHD-3β), while the most recent (ICHD-3) recommends the term prodrome [3, 8‐10]. In an editorial, the incoming Chairman of the Classification Committee, which on behalf of the International Headache Society is responsible for ICHD, expressed a personal opinion that premonitory symptoms should be reinstated as the preferred term in future iterations of ICHD [11]. Although this remains an unresolved issue, the two terms are for practical purposes synonymous, describing a symptomatic phase defined as above [3, 8‐10]. Since premonitory symptoms has been more widely used in the literature, we used this term for the purposes of this review.

To describe the percentage of people with migraine who experience premonitory symptoms, we used the terms prevalence when referring to data from population-based samples and relative frequency when describing data from clinic-based samples.

We followed the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) checklist [12], and registered the study protocol in PROSPERO (CRD42021255339).

We systematically searched PubMed and Embase from database inception until 31^st May 2022 for observational studies reporting the prevalence or relative frequency of one or more premonitory symptoms among people with migraine. The search string was “migraine AND (premonitory OR prodromal OR prodrome)”. After removal of duplicates, two of us (AKE and AI), independently, first screened titles and abstracts for relevance, then reviewed the retrieved full texts for eligibility based on pre-defined inclusion and exclusion criteria (Table 1). Eligible studies varied widely in the method of assessing premonitory features and in the extent to which an operational definition was provided. The reference lists of retrieved publications were also searched to identify other eligible studies. Final study selection was determined by consensus between AKE and AI.

Two investigators (AKE and RHC) independently extracted data from all studies according to a pre-defined set of variables (Tables 2 and 3), afterwards reaching consensus between them.

Any discrepancies during the processes of study selection and data extraction were resolved with the assistance of a third investigator (HA).

Two of us (AKE and RHC) independently assessed risk of bias using the Joanna Briggs Institute Critical Appraisal Instrument for Studies Reporting Prevalence Data (Supplementary Table 2) [29]. The instrument contains nine items: (1) Was the sample frame appropriate to address the target population? (2) Were study participants sampled in an appropriate way? (3) Was the sample size adequate? (4) Were the study participants and the setting described in detail? (5) Was the data analysis conducted with sufficient coverage of the identified sample? (6) Were valid methods used for identification of the condition? (7) Was the condition measured in a standard, reliable way for all of the participants? (8) Was there an appropriate statistical analysis? (9) Was the response rate adequate? If not, was the low response rate managed appropriately? Higher total scores indicated lower risk of study bias. Studies were categorized according to the percentage of yes answers as high risk (≤ 49%), moderate risk (50%-69%) or low risk (≥ 70%) [29].

We characterised studies as clinic-based or population-based, and separately analysed data from each. We performed a random-effects meta-analysis, which accounted for between-study heterogeneity and calculated pooled prevalence using the inverse variance method on logit transformed data. Between-study variance was calculated using the restricted maximum likelihood method. The I² statistic was used to assess between-study heterogeneity. I² signifies the amount of variation between studies that can be attributed to study heterogeneity rather than chance. Values ≥ 75% indicate considerable heterogeneity and, therefore, uncertainty surrounding pooled estimates. Meta-analysis was performed only when three or more studies reporting relevant outcomes with a total sample size of N ≥ 100 subjects were available. The limited data quantity could not support meaningful meta regressions or funnel plots. All statistical analyses were performed with R version 3.5.2 using the “meta” and “metafor” packages.

Overall prevalence was 29% (95% CI: 8–63; I² = 99%; N = 1,638) in population-based studies [13‐16] and relative frequency in clinic-based studies was 66% (95% CI: 45–82; I² = 99%; N= 10,107) [17‐28]. Of the 29 studies included in the qualitative analysis, five exclusively reported pediatric data and were therefore not included in the meta-analysis. Two clinic-based studies estimated overall relative frequency of premonitory symptoms in children with migraine: 42% in one (N = 176) [32] and 67% in the other (N= 103) [33]. One clinic-based study reported relative frequency in adolescents with migraine (N= 19), with follow-up assessments after two years [34]. Premonitory symptoms were reported by three of 19 subjects (17%) at baseline and by nine of 18 (50%) after two years.

Between them, the studies reported a total of 96 specific individual premonitory symptoms. Data were too few from population-based samples, but sufficient to support meta-analysis of 11 of these in clinic-based populations (in order of frequency: fatigue [49%], neck stiffness [46%], mood change [37%], concentration difficulties [30%], nausea [29%], photophobia [29%], phonophobia [26%], yawning [22%], depressive symptoms [19%], irritability [16%], food craving [11%]) (Table 3). Numbers of studies contributing to each analysis were low, and I²-values were high (> 85%) for all except concentration difficulties (17%) (Table 3).

Five studies recorded non-headache symptoms during premonitory, headache and postdromal phases [18, 22, 35‐37], the last defined in ICHD-3 as occurring in association with a migraine attack but after (up to 48 h) resolution of the headache [3]. All five found that non-headache symptoms commonly reported as premonitory, such as yawning and fatigue, were equally common during the headache and postdromal phases.

Of the 29 studies, 15 assessed prevalence or relative frequency of premonitory symptoms as the primary outcome [17, 19, 21‐25, 28, 31‐33, 35, 37‐39]. Nine of the 22 clinic-based studies collected data prospectively [17, 18, 22, 23, 26, 27, 35, 36, 39], while 14 [19‐21, 24, 25, 28, 30‐34, 37, 38, 40], and all six population-based studies [13‐16, 41, 42], acquired data retrospectively.

The studies applied different definitions of premonitory symptoms. Six studies used the criteria offered by the ICHD iteration at the time of conduct [17, 21, 28, 33, 37, 39]. Five other studies applied criteria otherwise in accordance with ICHD but specifying time intervals that were shorter or longer than the 48 h interval specified by ICHD-3 [15, 16, 25, 27, 32]. Seven studies used definitions that were not consistent with ICHD [18, 22‐24, 35, 38, 40]. Eleven studies provided no definitions [13, 14, 19, 20, 26, 30, 31, 34, 36, 41, 42].

Data were collected by face-to-face interview in 12 studies [13‐16, 20, 28, 31‐33, 36, 38, 42], by self-administered questionnaires in nine [21, 22, 24, 25, 30, 37, 39‐41], by subject-completed diaries in four [17, 23, 27, 35] and by combinations of these methods in four [18, 19, 26, 34]. Eighteen studies used pre-defined lists of 2–29 (from the total of 96) premonitory symptoms [15‐19, 21‐28, 31‐33, 35, 41], four asked specifically about a single specified symptom [36, 37, 39, 40], and seven did not report how they elicited premonitory symptoms. Seven studies using pre-defined lists also included the option of free recall [15‐17, 23, 28, 35, 41].

We assessed risk of bias as high in 20 studies [14, 17‐20, 22, 23, 25‐27, 30‐36, 38, 39], moderate in seven [15, 16, 21, 24, 28, 37, 42], and low in two [13, 41] (Supplementary Table 3). The majority of studies (24 of 28) used sampling frames that did not appropriately address the target population (Item 1): 23 clinic-based studies focused on migraine patients in the limited context of specialized or non-specialized headache clinics, while one, although population-based, included only specific subgroups of people with migraine. Twenty-three studies recruited participants inappropriately by convenience sampling, judgmental sampling or snowball sampling (Item 2). Sample size appeared inadequate (< N = 300) in 15 studies (Item 3). Subjects were not well characterized in 28 studies, with missing data relating to gender, age, proportions with and without aura, monthly headache days and/or monthly migraine days (Item 4). None of the studies used validated instruments to assess premonitory symptoms (Item 6). Five of nine longitudinal studies had high dropout proportions (> 15%); four others did not report dropouts (Item 9). Ten of 20 cross-sectional studies had responder proportions between 60% and 97.4%, while ten did not report responder proportions (Item 9).

The strength of this study lies in the systematic literature review and meta-analyses following standard (PRISMA) methodological guidelines. The limitations were not in the study itself but in the data. There was considerable between-study heterogeneity largely due to varying and often questionable methodologies. The majority of data came from selected (clinic-based) populations, with any biases this might have introduced (evidence of bias is seen in the different prevalence/relative frequency estimates: 29% population-based and 66% clinic-based). There were sparse data from a limited number of studies (and/or small sample sizes) for some of the meta-analyses and too few data to instigate a meta-analysis on pediatric studies. There were also too few data to permit us to perform meta-regression, which might have clarified the effects on study heterogeneity of different variables such as definition of premonitory symptoms and assessment methods. There were too few data to make a funnel plot to check for publication bias.

Since our findings represent the whole of the available evidence, the key question they generate is whether, in their objectivity or their totality, they confirm the existence of a premonitory phase as a distinct phase of the migraine attack. We are not at all sure they do. This is not to deny that the phase exists: it is an absence of evidence, not evidence of absence. There is work to be done, in five directions.

First, research must address the methodological shortcomings that are evident in the literature, and promote standardisation to make future studies more comparable. Above all, the field needs an operational consensus definition of the premonitory phase. Ultimately, ICHD should be the source of this definition, but our findings suggest that the current definition [3] needs revision and perhaps further specification. Revision requires more and better empirical data than currently exist. Studies designed to acquire these data must carefully consider how best to elicit premonitory symptoms objectively and reliably. Lists of predefined symptoms prompt recall, and are likely to result in higher estimates of prevalence or relative frequency, but they may also lead to false-positive symptom reporting through yea-saying, a well-known phenomenon in survey research [46]. We recommend beginning with open-ended questions, to be followed by lists that include some dummy (highly unfeasible) response options. Ideally, methodological guidelines agreed by expert consensus are needed.

Secondly, studies should be conducted in the population of interest. Clinic-based studies are feasible and of interest to clinicians. Generalizability from speciality care to primary care or from diagnosed patients in primary care to the general migraine population is hazardous. If the goal is to characterize premonitory features in unselected people with migraine, population studies are required. Population-based studies can establish the prevalence of premonitory symptoms and characterize their frequency, intensity and duration. There is, in particular, a need for population-based studies investigating premonitory symptoms in children; all five pediatric studies included in this review were retrospective, clinic-based studies with small sample sizes. To establish premonitory specificity, it is also necessary to enquire more closely into the prevalence (and relative frequency) of non-headache symptoms during all migraine phases, including the interictal period. This probably requires prospective studies with contemporaneous diary recordings. Studies that ask participants to recall what might have been premonitory symptoms after they have entered the headache phase, as did the majority of the studies in our review, invite recall error and are likely to introduce bias by reverse causality attribution [7, 47]. Prospective diaries can eliminate these, but only so long as they do not allow post-hoc data entry (or amendment) [43]. Unfortunately, this comes at a price: symptoms might be missed because they are not immediately recognised, and because the demands of daily (or more frequent) data entry are onerous, and likely to be neglected. Certain character traits are needed to meet these demands conscientiously, and this in itself introduces bias. The expert consensus guidelines called for above should take a view on this. Prospective diaries support collection of large amounts of data, and examination of multiple factors and any associations between them [7], but there is little benefit if the data are misleadingly unrepresentative.

Thirdly, are premonitory symptoms pathognomic of migraine? To answer this requires capturing pre-headache symptoms in people with migraine and tension-type headache (TTH), and determining which are optimally discriminative between these disorders. One of the included population-based study found that premonitory symptoms were no more frequent in people with migraine compared to those with TTH: the most frequent (low spirits and tiredness) being equally so among both groups [15]. If premonitory symptoms are confirmed as a feature of migraine, enquiry can proceed into how they relate to attack initiation. Are they post-onset, signalling the first beginnings of a migraine attack – a true premonitory phase, or are they pre-onset, epiphenomenal symptoms of a state in which the threshold to attack initiation is lowered?

Fourthly, what are the hallmarks of premonitory symptoms? While some might argue that they should be specific to the pre-headache period, others might argue that the essential hallmark is association, at within-person level, with an increased probability of headache over a defined succeeding period. To implement this definition, prospective diary studies are required, so that within-person risk of headache can be calculated over periods that do and do not follow the recording of a particular symptom or symptom complex. This is the approach that would be most useful to support either biological observation or studies of pre-emptive therapy. Limited work has focused on symptoms that predict future attacks [48].

Finally, population-based studies might enquire into how much premonitory symptoms contribute to migraine-attributed burden. With nothing known of this, it would be an important line of enquiry. Burden-of-migraine studies, including the Global Burden of Disease studies, which rank migraine second among causes of lost health worldwide, have focused entirely on the headache phase of migraine [49, 50].