This study was approved by the Ethics Committee of Tangdu Hospital (TDLL-2015133). Informed consent was not required for this study as it exclusively relied on the analysis of observational data obtained from the patient information administration and registration system. Totally 736 patients diagnosed with migraine were retrieved, from January 2002 to January 2023. The diagnosis was made based on the International Classification of Headache Disorders, 3rd edition (ICHD-3).

The inclusion criteria: patients with a confirmed diagnosis of migraine (including recurrent headache attacks lasting 4–72 h; headache is unilateral location; pulsating quality; moderate or severe pain intensity; aggravation by or causing avoidance of routine physical activity; association with nausea and/or photophobia and phonophobia); patients with available data on gender, age, trigger factor, clinical symptoms, comorbidity, and other relevant suffering performance. Exclusion criteria include: other primary headache disorders (e.g., cluster headache, tension-type headache, medication-overuse headache, post-traumatic headache, etc.), neurological diseases, psychiatric disorders, and rheumatic diseases.

A total of 99 individuals were excluded, among whom 27 were due to repeated admissions, 26 due to incompleted medical history records, and 46 because migraine was listed as a comorbid diagnosis (12 participants for malignant/metastatic tumor, 26 participants for cerebral hemorrhage sequelae, and 8 participants for medication overuse headache).

Ultimately, 637 patients were included in this analysis. The following demographic data and headache characteristics were obtained: age, gender, headache location (including left, right, bilateral side, or others), and the nature of pain (pulsating, bursting, stabbing, or others).

Moreover, clinical characteristics were collected: prodromal symptoms including blurred vision and dizziness, trigger factors such as mood changes, fatigue, influenza, and menstruation. Suffering symptoms, photophobia, phonophobia, nausea, and vomiting, were also included. All those symptoms were self-reported by the migraine patients and confirmed during consultation with a neurologist. The comorbidities included hypertension, type 2 diabetes mellitus (diabetes), patent foramen ovale (PFO), anxiety/depression, and cerebrovascular disease (including lacuna infarction, white matter hyperintensities, and lacunar ischemic stroke). The anxiety/depression levels of the patients were assessed by the Self-rating Depression Scales (SDS) (17) and Self-rating Anxiety Scales (SAS) (18). PFO was diagnosed using contrast echocardiography.

The dataset of migraine patients was randomly divided into training and validation cohorts at a ratio of 7:3, and the variables were compared. Normally data were listed as mean ± standard deviation (SD), while non-normal data were presented as median (interquartile ranges). In the univariate analysis, the chi-square test or Fisher’s exact test was used to analyze the categorical variables. Comparisons were using the Student’s t-test or Mann–Whitney U test to examine the continuous variables. To reduce the effect of multicollinearity on the regression results, the least absolute shrinkage and selection operator (LASSO) logistic regression (19) analysis was used for multivariate analysis to screen the independent risk factors, and the logistic regression model was used to establish a predictive nomogram for MWoA in the training cohort. The performance of the nomogram was assessed using the receiver operating characteristic (ROC) curve, with the area under curve (AUC) ranging from 0.5 (no discriminant) to 1 (complete discriminant). The Hosmer-Lemeshow goodness of fit test in the multiple logistic regression was performed to assess the model calibration and the calibration plot was plotted. A decision curve analysis (DCA) was also performed to determine the net benefit threshold of prediction. There were missing data for patients’ headache characteristics (like headache location, pain nature), triggers for headache attacks, and comorbidities. In the logistic regression model, only the subjects with complete data in all variables were considered. The multiple imputations were used to assess the sensitivity of results to missing values (20). Results with a p value of <0.05 were considered significant. All statistical analyses were performed using the R software (Version 4.2.2).

Of the included 637 migraine patients, 477 (46.62 ± 15.64) were females and 160 (39.78 ± 19.53) were males. There were no significant differences observed in the distribution of different age groups among the training test cohort (N = 446) and internal validation test cohort (N = 191) (Table 1).

The majority age group of participants was the 45–65 years category, representing 46.5% of the overall individuals, with a similar proportion in the two test cohorts. The females comprised approximately 74.9% of the total sample, and the distribution also remained across cohorts.

The attacking headache of migraine was located in bilateral (36.73%) with throbbing nature (37.68%), a similar pattern across cohorts. Fatigue (14.76%) emerged as the predominant trigger factor. The higher proportion of prodromal symptoms were dizziness (34.85%) and blurred vision (20.41%). Nausea (48.80%) and vomiting (32.03%), although commonly suffered symptoms of migraine, and no substantial significant difference across cohorts was detected (Table 1).

The prevalence of comorbid conditions such as PFO, hypertension, diabetes, and cerebrovascular disease was consistent among the cohorts, with no significant differences. Overall, the baseline characteristics of the study population were largely consistent across cohorts, providing a solid foundation for further predictive analyses (Table 1).

There was no significant variance inflation factor (VIF) among variables, there were no outliers, and the included variables met the linearity assumption (Supplementary Table S2). Multiple imputations for missing data generated similar results.

Using LASSO regression analysis performed in the training cohort, 6 potential predictors were included from the candidate variables in the original model, Age, Gender, Location of headache, Headache nature, Mood changes, Fatigue, Influenza, Blurred vision, Dizziness, Photophobia, Phonophobia, Nausea, Vomiting, Hypertension, Diabetes, PFO, Anxiety/Depression, and Cerebrovascular disease (Figure 1).

The coefficients of included predictors were estimated, blurred vision (−0.33), dizziness (−1.31), nausea (−0.09), vomiting (−0.07), PFO (0.17), and anxiety/depression (−0.03). A cross-validated error plot of the LASSO regression model was also exhibited. The most regularized and parsimonious model, with a cross-validated error within one standard error of the minimum, conclusively included 6 variables. The receiver operating characteristic curves (ROC) were yielded for the abovementioned variables’ area under curve (AUC) values greater than 0.5 and the prediction model was derived from the multivariate logistic regression for training and validation cohort (Figure 2; Supplementary Table S1).

The nomogram and calibration plots among the different cohorts were plotted, which demonstrated a good correlation between the observed and predicted migraine subtype (Figure 3). Findings indicated that the initial nomogram remained applicable for the validation sets, with the calibration curve of the model closely approximating the ideal curve (Figure 4). The Hosmer-Lemeshow goodness fit showed a p-value of 0.598 for the training cohort and 0.179 for the validation cohort. These results suggest a high level of consistency between predicted outcomes and actual observations.

Nomogram for predicting the probability of MWoA. The presence or absence of each clinical characteristic indicates a certain number of points. For each characteristic, absence is assigned 0 points. The presence of characteristic is generated using R based on the results of LASSO analysis. The points for each characteristic are summed together to generate a total points score.

Further multivariate logistic analyses were carried out in the training cohort. We found that PFO (OR = 2.30, p = 0.01), blurred vision (OR = 0.40, p = 0.001), dizziness (OR = 0.16, p < 0.01), and anxiety/depression (OR = 0.41, p = 0.02) was the influencing factor for the MWoA. Symptoms of nausea (OR = 0.79, p = 0.43), and vomiting (OR = 0.64, p = 0.17) did not enhance the risk of the suffering of MWoA (Table 2).

According to the diagnostic criteria of the ICHD-3, nausea and vomiting are essential symptoms for the diagnosis of migraine. Additionally, after removing these two variables, the AUC value of the model decreases (the AUC values for the training cohort and validation cohort were 0.773 and 0.826, respectively, Supplementary Figure S1). Meanwhile, these two variables exhibiting a high prevalence among migraine patients were included in the final model to intensify its clinical practical guidance significance.

Variables (such as blurred vision, dizziness, nausea, vomiting, and anxiety/depression) with a negative β value can be considered protective predictive factors for the diagnosis of MWoA, whereas PFO is a risk predictive factor for MWoA.

The DCA curve of the nomogram indicated the chance of substantial divergence in the predictive accuracy of the model when clinicians encounter imperfections during the utilization of the nomogram for diagnostic deliberation and the diagnostic decision-making processes. The present study demonstrated that the nomogram confers considerable advantages in terms of clinical utility, as evidenced by its favorable performance on the DCA curve (Figure 5).

The relationship between clinical characteristics and the risk of migraine was initially investigated in this study, and their predictive for MWoA was assessed subsequently. The results not only validated the predictive significance of PFO, blurred vision, dizziness, and anxiety/depression for MWoA, but also indicated a substantial association between PFO and MWoA, as well as a certain predictive capacity.

The PFO is a remnant of fetal circulation and is also the most common congenital cardiac anomaly in adult populations (23). Multiple studies suggest that migraine is more prevalent in subjects with PFO and vice versa (24). Several studies found that the incidence of PFO in migraine was 14.6–66.5% (25). In turn, in the population with PFO, the incidence of migraine was 9.13–51.7% (24). Some studies have shown that there is a stronger relationship between migraine with aura and PFO, the incidence of PFO is 46.3–88% in migraine patients with aura compared with 16.2–34.9% in migraine patients without aura (24). It has also been found that the atypical aura group was 79.2% vs.46.3% in the typical aura group (26). In China, a community-based cross-sectional study pointed to a strong association between PFO and MWoA, especially when the shunt is large (27). This difference might be due to racial genetic or regional disparities.

Therefore, we suggest that the pathophysiological relationship between PFO and aura in migraine patients needs to be further explored. However, the likelihood of benefit from PFO closure would appear to be increased, especially for migraine sufferers who have failed multiple pharmacological interventions (28), by cessation of migraine headaches or reducing migraine attacks and migraine days (29).

PFO can also serve as a significant sign for early classification and diagnosis of migraine patients. It can be considered that patients with PFO have a higher probability of suffering from migraine, and conversely, patients with migraine are more likely to have a PFO.

Anxiety/depression has a significant association with migraine allodynia (30), and migraine patients were more likely to be females who reported higher levels of current anxiety symptoms (31). Epidemiological studies have shown that patients with migraine were three times more likely to suffer from depression (1) and four to fivefold increase in the risk of anxiety (32) than the general population.

In recent studies, anxiety/depression has been confirmed a close correlation with migraine (33). Patients with migraine had worse depression and anxiety than those without (34). Depression and anxiety were the predictive factors of beliefs related to constancy of pain (35). Depression, headache features (higher headache pain intensity, more headache days per month), and no pharmacologic treatment factors (not using preventive migraine medications) were significant predictors of inadequate 2-h pain-free (36). There is a higher risk of transformation that migraine with anxiety/depression to chronic migraine, and risks are more pronounced in migraine with aura than in MWoA (2). Consistent with previous studies, this study demonstrated that MWoA had a lower risk of developing anxiety/depression than other patients.

The prodromal phase symptoms can be in some situations confused with migraine aura (37), and the description is not clear enough especially for the visual symptoms (38). It can be described as “foggy vision,” “dimness,” blurred vision for near or far objects, and blind or black spots. The prodromal symptoms of migraine, such as sensitivity to light or noise, neck stiffness, fatigue, and difficulty concentrating, had been reported to be prevalent in migraine patients through the studies of questionnaires, clinician interviews, diaries, or retrospective recall (39, 40). From 122 migraine with aura patients shown that the most visual symptom was blurred vision (41). Blurred vision accounted for 17.9% of participants of chronic migraine (42). Friedman and Evans have proposed that blurred vision may be a symptom of autonomic dysfunction due to an imbalance in sympathetic and parasympathetic signaling (43).

Along with dizziness, migraine was also characterized by symptoms such as phonophobia, motion intolerance, nausea, vomiting (44), and photophobia (45). In fact, up to 15.5% of visits to general healthcare settings were related to concerns about dizziness (46). Interestingly, in some migraine patients, the dizziness and vomiting can be more debilitating than the headache itself (47).

Meanwhile, MWoA had a lower risk of suffering prodromal symptoms like blurred vision, dizziness, and common symptoms of nausea and vomiting. This indicates, from one perspective, that patients with MWoA tend to have few other positive physical signs beyond their headache symptoms, necessitating heightened attention during clinical differential diagnosis. For migraine, early intervention may be the most effective acute treatment strategy (8). It also suggests that when treating and evaluating treatment efficacy, particular emphasis should be placed on the alleviation of headache symptoms.

This predictive model ultimately utilized PFO, blurred vision, dizziness, nausea, vomiting, and anxiety/depression to identify MWoA, which has a high ROC value. Meanwhile, when LASSO regression was used to select variables, variables such as gender, age, occupation, and the headache nature were excluded, making the model more convenient.

The AUC values for the training and validation cohort were 0.791 and 0.828, respectively, indicating that this nomogram has good accuracy and stability. The nomogram is relatively easy to use due to the fewer number of included variables, and the operation is also easy to master. This model can obtain a total score of variables and the probability of developing MWoA, thus helping clinicians provide more beneficial advice to patients. For example, if a migraine has a PFO and does not experience symptoms such as blurred vision, dizziness, nausea, vomiting, or anxiety/depression, their total score would be approximately 270, corresponding to a probability of about 90% for MWoA. The developed nomogram offers several clinical implications. Firstly, it provides a quantitative tool for clinicians to predict MWoA more accurately than traditional methods, aiding in better risk stratification for comorbidity PFO. Moreover, early identification of high-risk individuals through this nomogram can lead to timely interventions, potentially reducing morbidity and mortality.

Our study has several limitations that should be acknowledged. The single-center retrospective study was based on patients from a hospital in Northwest China, which may not be representative of the wider population. Furthermore, there might be potential unmeasured confounders that were not included in our model. External validation in diverse populations will be essential to confirm the generalizability of our findings.