Introduction

Migraine prevention has advanced thanks to the introduction of new target-driven treatments antagonising the calcitonin gene-related peptide (CGRP) or its receptor.1 Specifically, galcanezumab, fremanezumab and eptinezumab are monoclonal antibodies targeting CGRP while erenumab targets the CGRP receptor (anti-CGRP MAbs). They have demonstrated their efficacy and tolerability in clinical trials2 and real-world studies3–8 for episodic (EM) and chronic migraine (CM) and are currently available in Europe.

Since anti-CGRP MAbs introduction, several questions have been raised about the treatment response to these drugs. First, CGRP is differentially modulated according to sex9 and likely to age,10 which may result in different response rates based on these demographic characteristics. Second, patients with EM can evolve to CM throughout their life11 12 and treatment responses to anti-CGRP MAbs can be affected by the severity of the disease and headache frequency. Lastly, from the current use of these drugs, around 15%–25% of patients emerged as non-responders (<30% reduction in monthly headache days—MHD)13 14 and another 15%–25% as excellent responders (≥75% reduction in MHD),3 15 16 but predictors of response or non-response (NR) are currently lacking.

All these questions have potentially relevant implications in terms of patient management and drug placement both in clinical practice and for regulatory agencies. Yet, they have been only marginally addressed to date.15–19 One of the reasons is that real-world studies, although particularly valuable to capture the clinical setting where anti-CGRP MAbs are used, have often limited sample size,20 predominantly because they are usually monocentric with precise prescription criteria. Clinical trials, on the other side, often have very specific inclusion criteria for participants, limiting the generalisability of their output to the broader spectrum of patients seen in daily practice.21 22

Thus, our study aimed to investigate factors influencing good and excellent response (ER) to anti-CGRP MAbs at 6 months in a real-world large European cohort of migraine patients.

Methods

This is a prospective multicentre real-world study. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline has been followed.23 To participate, all centres were required to have an available prospective dataset of patients fulfilling the International Classification of Headache Disorders, third edition (ICHD-3) criteria for migraine,24 either EM or CM, and treated with anti-CGRP MAbs since March 2018, according to their own country prescription and reimbursement policy (online supplemental table 1). Although each country presented slightly different prescription criteria, our cohort was mainly composed of resistant patients with migraine who had failed multiple previous preventive treatments.25 We included patients with concomitant migraine preventive medications and patients with medication overuse.

Each centre provided the following data to the multicentre study dataset: baseline demographics (age, sex and reproductive status), comorbidities (arterial hypertension, obesity, any cardiovascular disease, anxiety and depression, which were collected from medical records), migraine characteristics (aura, allodynia, accompanying symptoms, unilateral pain, unilateral cranial autonomic symptoms or signs—as defined by ICHD-3 for trigeminal autonomic cephalalgias, ipsilateral to the headache and present in the majority of patient’s migraine attack—time since migraine onset and chronification, if applicable), previously failed migraine preventive treatments, concomitant migraine preventive medications, date and type of anti-CGRP MAbs prescribed (erenumab 70 mg or 140 mg monthly, galcanezumab 120 mg monthly+240 mg loading dose, fremanezumab 225 mg monthly or fremanezumab 675 mg quarterly). Outcome variables used were MHD, monthly migraine days (MMDs) and monthly acute medication days (MAMDs). These data were collected through patients’ prospective headache diaries, either in paper or electronic depending on the centre, at baseline (M0), 3 months (M3) and 6 months (M6). The patient-reported outcomes questionnaires included in this study were the Migraine Disability Assessment (MIDAS)26 and the Headache Impact Test (HIT-6).27 We excluded patients with missing age, sex and MHD data at baseline or at follow-up.

Response to anti-CGRP MAbs was categorised according to the mean reduction in MHD at 6 months from baseline: NR (NR—MHD reduction <30%), partial response (PR—MHD reduction 30%–49%), good response (GR—MHD reduction ≥50%) and ER (ER—MHD reduction ≥75%).

We then searched the variables that were independently associated with GR and ER at 6 months.

Statistical analysis

We reported nominal (categorical) variables as frequencies (percentages), whereas the median and IQR were reported for continuous variables. We checked the normality assumption of continuous variables through visual methods (Q-Q plots). In the final dataset, we detected a rate of missing ranging from 4.6% to 44.3% in some variables, showing a missingness rate≥25% in MIDAS (44.3%), allodynia (40.8%), depression (26.3%) and unilateral cranial autonomic symptoms (25.7%) (online supplemental figure 1). We used random forest imputations in order to estimate these values according to basal MHD, age, sex and reproductive status using the multivariate imputation via chained equations package from R (V.3.16.0).28 Then, we tested in a sensitivity analysis the consistency of the main results reported in the raw dataset (without data imputation). Statistical significance between treatment responder’s subgroups was performed with Fisher’s exact test when comparing categorical variables, independent t-test was used for normal quantitative variables and the Wilcoxon rank-sum test for not-normally distributed quantitative variables.

Then, a generalised mixed-effect regression model (GLMM) was estimated for a binary outcome variable (responder vs non-responder) in order to identify variables associated independently with treatment response rate. GLMM was, therefore, used for both GR and ER analysis. All participants, no matter if they continued or discontinued the anti-CGRP MAbs because of lack of efficacy, were included in the data modelling to avoid obtaining biased results due to treatment discontinuation. First, data were split into two subsets: training set (80% initial data) and test set (20% initial data). Then, a full GLMM was fitted within training dataset, adjusted by fixed-effect covariates (concomitant medication) and random effects (patients, headache centres and type of anti-CGRP-MAb). GLMM parameter estimation was performed using restricted maximum likelihood estimation. The best-fitting model was obtained according to the minimum corrected Akaike information criterion (AICc) and likelihood ratio tests were performed to ensure that the best AIC model was better than the full model. The final model was validated using repeated 10-fold cross-validation with 3 repetitions and model’s accuracy was evaluated in the test dataset, computing the corresponding confusion matrix and ROC curve.

Models were fitted using R package glmmTMB V.1.1.7, variable selection was obtained using R package MuMIn V.1.1.7. Variance inflation factors for all the parameters were computed in order to estimate how much the variance of an estimated regression coefficient is inflated due to correlated variables so that we could avoid an overfitting problem in the final models. The analysis of the deviance table of the model’s main effect was performed and main effect plots were plotted using the R sjPlot package V.2.8.14. Variable importance of each predictor was estimated in terms of the differences in AICc between the full model and the model without each predictor.

We did not conduct a statistical power calculation prior to the study because the sample size was based on available data. P values presented are for a two-tailed test and we considered p values <0.05 as statistically significant. Due to the exploratory nature of the present study, all p values were adjusted by applying the false discovery rate (FDR, Benjamini-Hochberg procedure). All analyses were done by using R V.4.3.0.

ResultsCohort description

35 European hospitals from 7 countries participated (Spain, Italy, Portugal, the UK, Germany, Norway and Poland) with a total population of 6281 migraine patients who were started on anti-CGRP MAbs. We excluded 7.4% (463/6281) of patients during the data quality check and we finally included 5818 patients: 98.7% of data available at M3 (5742/5818) and 85.3% of data at M6 (4963/5818). The flow chart of the study cohort is shown in figure 1.

Figure 1

Flow chart of the study cohort. We included patients who started anti-CGRP MAbs from March 2018 to May 2023 who had available data at follow-up during the first 6 months of treatment. anti-CGRP MAb, anti-CGRP monoclonal antibody; FUP, follow-up; M0, baseline (month 0); M3, month 3; M6, month 6; MHD, monthly headache days.

At baseline, of the 5818 patients, 82.3% (4786/5818) were female and the median age was 48.0 (40.0–55.0) years. The most represented countries were Spain (63.9%; 3719/5818) and Italy (22.5%; 1310/5818). Anxiety (34.7%; 2017/5818) and depression (29.6%; 1724/5818) were the most frequent comorbidities. 72.2% (4198/5818) had a baseline diagnosis of CM. The median MHD, MMD and MAMD were 20.0 (14.0–28.0) days/months, 15.0 (10.0–20.0) days/months and 15.0 (10.0–25.0) days/months, respectively. The median number of preventive treatment classes failed was 4.0 (3.0, 5.0). Table 1 shows all collected baseline variables.

Table 1

Baseline characteristics of the study cohort

Effectiveness and tolerability

At month 6, we observed a median reduction in MHD of −9.0 (−15.0, –3.0) days/month. The response rates were NR 30.3% (1503/4963), PR 13.2% (656/4963), GR 56.5% (2804/4963) and ER 26.7% (1324/4963).

Treatment discontinuation occurred in 8.0% (462/5742) at M3 and in 7.4% (369/4963) at M6, mainly because of lack of efficacy (M3: 84.6%, 391/462; M6: 90.8%, 335/369). Discontinuation due to lack of tolerability at any time point occurred in 1.6% (91/5742). The presence of any adverse event (AE) was reported in 22.7% (1304/5742) at M3 and in 18.5% (920/4963) at M6, classified as mild. Main AEs reported were constipation (M3: 52.8%, 688/1304; M6: 54.1%, 498/920) and injection site reaction (M3: 14.0%, 183/1304; M6: 15.7%, 144/920). No patient had serious major safety concerns.

Variables associated with a GR at 6 months

The statistically significant variables of the univariate analysis, comparing patients with and without GR (≥50% vs <50% MHD reduction), are shown in online supplemental table 2.

In the GLMM, statistically significantly independent variables associated with ≥50% MHD reduction were older age (1.08 (95% CI 1.02 to 1.15), p=0.016), the presence of unilateral pain (1.39 (95% CI 1.21 to 1.60), p<0.001), the absence of depression (0.840 (95% CI 0.731 to 0.966), p=0.014), less failure to onabotulinumtoxinA (BTX-A) (0.786 (95% CI 0.676 to 0.913), p=0.002) but more failure to beta-blockers (1.17 (95% CI 1.02 to 1.35), p=0.031), less concomitant oral medication (0.873 (95% CI 0.766 to 0.995), p=0.042), less MMD (0.923 (95% CI 0.862 to 0.989), p=0.023) and lower MIDAS at baseline (0.874 (95% CI 0.819 to 0.932), p<0.001) (table 2 and figure 2B–I). Figure 2A shows the importance of these variables within the model. The main results were reproducible in the sensitivity analysis (online supplemental table 3).

Table 2

Mixed-effects logistic regression model for the prediction of MHD reduction ≥50% and ≥75% after 24 weeks of anti-CGRP monoclonal antibodies treatment

Figure 2

Relative variable importance within the model (A) and fixed-effect plots (B–I) of the final mixed-effects logistic regression model for the prediction of MHD reduction ≥50% after 24 weeks of anti-CGRP monoclonal antibodies treatment. Relative variable importance of predictors was assessed by observing how much the model’s AIC changes when each predictor is removed individually. It provides insights into which predictors have a more significant impact on the model’s fit. AIC, Akaike information criterion; BB fail., failure to beta-blockers; Dep: depression; est. prob., estimated probability; MIDAS, migraine disability assessment; onabotA fail., failure to onabotulinum toxin A; MMD, monthly migraine days; Oral conc., concomitant oral medication; RR, response rate; MHD, monthly headache day; Unil. Pain, unilateral pain.

The final fitted model presented an accuracy (95% CI) of 0.644 (0.630 to 0.658) and Area under the ROC Curve (AUC) of 0.692 (0.677 to 0.707) in the training set and an accuracy of 0.613 (0.583 to 0.641) and AUC of 0.648 (0.616 to 0.680) in the test set (online supplemental figure 2A).

To better understand the influence of age on the probability of treatment response, we also conducted a sensitivity analysis of age cut-offs, where we found that participants aged 50 years or older had a higher probability of achieving good treatment response (51–55 years: 1.32 (95% CI 1.06 to 1.65), p=0.014; >56 years: 1.27 (95% CI 1.04 to 1.56), p=0.019). We conducted the same sensitivity analysis for MMD and we observed that only participants with baseline MMD greater than 20 days had a lower probability of achieving good treatment response (0.762 (95% CI 0.605 to 0.961), p=0.021).

Variables associated with an ER at 6 months

The statistically significant variables from the univariate analysis, comparing non-responders (<30% reduction in MHD) with excellent responders (≥75% reduction in MHD) at month 6, are shown in online supplemental table 4.

In the GLMM, statistically significantly independent variables associated with ≥75% MHD reduction were older age (1.19 (95% CI 1.09 to 1.30); p<0.001), migraine diagnosis at baseline (CM) (1.91 (95% CI 1.50 to 2.43), p<0.001), the presence of unilateral pain (1.63 (95% CI 1.36 to 1.97), p<0.001), the absence of depression (0.705 (95% CI 0.584 to 0.850), p<0.001), less failure to BTX-A (0.699 (95% CI 0.568 to 0.861), p=0.001) and antihypertensive drugs (0.724 (95% CI 0.569 to 0.923), p=0.009), less concomitant oral medication (0.722 (95% CI 0.606 to 0.862), p<0.001), less MMD (0.767 (95% CI 0.690 to 0.852), p<0.001) and lower MIDAS at baseline (0.890 (95% CI 0.814 to 0.974), p=0.012) (table 2 and figure 3B–J). Variable importance is plotted in figure 3A. Main results were reproducible in the sensitivity analysis (online supplemental table 3).

Figure 3

Relative variable importance within the model (A) and fixed-effect plots (B–J) of the final mixed-effects logistic regression model for the prediction of MHD reduction ≥75% after 24 weeks of anti-CGRP monoclonal antibodies treatment. Relative variable importance of predictors was assessed by observing how much the model’s AIC changes when each predictor is removed individually. It provides insights into which predictors have a more significant impact on the model’s fit. AHD fail.: failure to antihypertensive drugs; MHD, monthly headache day; AIC, Akaike information criterion; Dep.: depression; DX: migraine diagnosis (EM, episodic migraine or CM, chronic migraine); est. prob.: estimated probability; MIDAS: migraine disability assessment; MMD: monthly migraine days; onabotA fail.: failure to onabotulinum toxin A; Oral conc.: concomitant with oral medication; Unil. Pain: unilateral pain.

The fitted model presented an accuracy (95% CI) of 0.685 (0.667 to 0.702) and AUC of 0.736 (0.717 to 0.755) in the training set and an accuracy of 0.657 (0.620 to 0.693) and AUC of 0.691 (0.651 to 0.731) in the test set (online supplemental figure 2B).

We conducted the same sensitivity analyses for age and MMD as we did for GR and we observed the same cut-offs for these variables also for excellent treatment response.

Discussion

At present, this is the largest real-world cohort of migraine patients treated with anti-CGRP MAbs. Baseline characteristics, effectiveness and tolerability are similar to the ones previously observed in other smaller real-world studies.3–5 7 8 However, our sample size allowed a more robust analysis of the factors influencing treatment response, thanks to the large representativeness of subgroups like men, older patients, non-responders, excellent responders, etc, usually very scarce in observational studies and clinical trials. We have obtained these results with consequently relevant implications for clinical practice.

First, treatment responses to anti-CGRP MAbs do not depend on sex. This is in line with a previous retrospective study with a 3-month follow-up29 and may suggest a second-tier role of sex-related CGRP modulation at least in terms of drug effectiveness. Reproductive status did not emerge as a relevant variable in our model either. Gender perspective in healthcare is important for more personalised treatments, but our study does not support, at present, that anti-CGRP MAbs should have a preferred use based on sex. Nevertheless, new studies should further look into this research question, also considering that for other anti-CGRP drugs, the gepants, when used for migraine acute treatment, sex differences in response to treatment may exist.30

Second, treatment responses to anti-CGRP MAbs depend on age. Interestingly, the older the patient, the higher is the likelihood of response. This is a novel finding that confirms recently published results indicating similar efficacy in patients ≥65 years old,31 32 underscoring the importance of age in relation to the biology of migraine and CGRP. The explanation behind this evidence is elusive at this time but may reflect the presence of age-dependent CGRP modulation, or a more relevant role of CGRP in migraine pathogenesis when other factors such as hormonal changes disappear.33 As a clinical implication, our study prompts to reconsider the current management of older migraine patients, often excluded in clinical trials or undertreated because of potential harm. Considering the longer life expectancy and age-related limitations of several oral preventive treatments, anti-CGRP MAbs appear to be a valuable option in older patients, this is starting from 50 years of age, in the absence of other contraindications.

Third, migraine frequency influences treatment responses: the higher the number of MMD, the lower the likelihood of good or ER. The same trend is observed for migraine disability (MIDAS): the higher the disability, the lower the probability of response. These results are supported by other studies reporting the negative impact of daily headache in anti-CGRP MAbs response.34 The finding regarding the higher probability of CM to have ER seems apparently in contrast with the above-described pattern of better outcome in less affected patients. The fact that EM patients in our cohort have high frequency, being similar to CM,35 and that the diagnosis is determined at a specific time point, no capturing patients’ cyclic migraine behaviour, makes this variable probably more fictitious and little informative, whereas what should be considered clinically meaningful is migraine frequency. The influence of migraine frequency in anti-CGRP MAbs responses is a fundamental point raised by this study. Although anti-CGRP MAbs were approved by EMA starting from four migraine days/month,36–38 their higher cost compared with other migraine preventive treatments led to prescription/reimbursement restrictive criteria in the majority of countries and makes, at present, these drugs available to a limited group of patients with higher migraine frequency and several other preventive treatment failures. Thanks to this multicentre collaborative effort, our study has the strength of including a broad range of baseline headache frequency, something that is lacking in the current literature. Our data corroborate the higher effectiveness of an early treatment with anti-CGRP MAbs, as shown by a negative linear correlation between MMD and probability of treatment response that, translated into clinical practice, should warn policy-makers that current reimbursement criteria, covering prescription to patients with high migraine frequencies, have the potential disadvantage of reducing the likelihood of response to treatment. In light of this finding, reimbursement criteria could be reconsidered, especially for those countries where only CM is reimbursed, as the likelihood of response is clearly reduced when patients have more than 20 MMD. In addition, it is important to consider the potential evolution from EM to CM,12 and the reduced therapeutic armamentarium for EM patients, this is, the lack of BTX-A approval for EM.

Finally, only considering clinical data, we were able to predict treatment response with a reasonable accuracy. The appropriate interpretation of treatment response comes from considering the combined effect of several variables simultaneously, rather than individually. For example, people who are young but also have high migraine frequency and disability will have less chances to respond than older individuals with less MMD. Our statistical models show that, among the combined set of variables that better predict good or ER, other factors also play an important role. Unilateral pain has emerged as a predictor of response in our study as well as in other observational studies14 16 39 40 and considering that it is one of the hallmarks of trigeminovascular activation,41 it suggests that patients with a well-defined migraine phenotype are more likely to respond. In addition, in the univariate analysis, we also found that cranial autonomic symptoms were associated with either good or ER, in line with previous studies,16 42 but eventually, this variable did not add any further information to our prediction model. Among comorbidities and previous treatment failures, having depressive symptoms and previous failure to BTX-A emerged as independent factors associated with less likelihood of anti-CGRP MAbs response, respectively. Migraine patients with depression seem to have higher levels of CGRP compared with those without, especially when they have higher migraine frequency,43 this may explain why this comorbidity influences treatment response, as observed also in other studies.44–46 Concerning BTX-A, recent studies have shown that mechanistically this drug also acts on CGRP release47 and may, in part, explain why failure to this treatment is a predictor of less likelihood of anti-CGRP MAbs response. Also, BTX-A is approved only for CM, which represents a severe form of the disease. The number of drug classes previously failed, on the contrary, did not add any relevant information in the models, indicating that patients respond to anti-CGRP MAbs regardless of the number of preventive treatments previously tried.16 48–50 The implication of having strong clinical predictors is twofold. In fact, selecting patients who are more likely to respond allows more precise migraine care and a reduction in healthcare costs. Overall, our work highlights the importance of properly phenotyping patients, especially through characterisation of migraine features that are able to guide clinicians in potentially detecting responders.51 Nevertheless, to achieve more precise predictions of anti-CGRP MAbs response in the future, we believe that more extensive, complete and standardised clinical data collection should be endeavoured in daily practice and integrated with non-clinical measures such as molecular and/or functional data.

This study has several limitations. First, because of the multicentric nature, differences in data collection across centres may have occurred; however, the strength of our study is to use of an adjusted GLMM model to minimise this issue. Moreover, we imputed missing clinical data based on patients’ demographics, (age, sex, MHD) under the assumption that the remaining clinical variables may depend on these factors. However, we conducted a sensitivity analysis to assess the robustness and consistency of our results using the raw dataset. We are also aware that, because of reimbursement criteria in most of the participating countries, our population mainly included high-frequency EM and CM patients with failure to several other preventive treatments, therefore, limiting the possibility of understanding the influence of low-frequency EM and treatment-naiveness on anti-CGRP MAbs responses and, therefore, generalising our results to the entire migraine population. However, these treatment-resistant patients represent the majority of those seeking medical care in headache clinics. In addition, we also recognise that other variables, not included in this study, may affect anti-CGRP MAbs responses, nevertheless our study aimed to focus on the most relevant and commonly collected data in clinical practice with the strength of counting on an ample study cohort. Finally, we are aware that at present predictors of treatment response are scarcely translatable to single individuals, in other terms, patients with some negative predictors may still have GR to the treatment and it would be inequitable to exclude patients from receiving it. However, our main study objective was not to predict treatment for single individuals, but rather to gain insights on potentially relevant aspects that influence treatment response at group levels, as a way to help improving migraine care.