Migraine is a highly disabling neurological disorder, with a prevalence exceeding that of diabetes, epilepsy, and asthma combined (1). It is more common in women than in men, with a global age-standardized prevalence of 18.9% in women and 9.8% in men (2). Estimates of the prevalence of migraine in Japan are slightly lower, ranging from 6.0 to 8.6% (3, 4). According to the Global Burden of Disease 2016 survey, it is the second leading cause of disability worldwide (1), particularly in women under the age of 50 years (2).

Despite its prevalence, migraines have only recently been recognized as an important public health concern (5). It was not included in the Global Burden of Diseases, Injuries, and Risk Factors studies before 2000 (2). Migraine research has historically been underfunded (6, 7) partially because it is an “invisible” disorder (8) and there are no obvious physical manifestations or clinical laboratory results associated with migraine. Consequently, this condition is often underestimated, underdiagnosed, and undertreated (9–12). However, its impact on productivity and quality of life is substantial (13). According to an employee population survey in Japan, the estimated annual economic loss due to presenteeism was USD 2,217 per person (5). Therefore, further research on migraine should be conducted in Japan to reduce this large patient burden.

The diagnosis of migraine is based on the symptoms reported by the patients. The diagnostic criteria for migraine, according to the International Classification of Headache Disorders, version 3 (ICHD-3), are attacks lasting for 4–72 h with any two of the following attributes: unilateral location, pulsating quality, moderate to severe pain intensity, and aggravation by physical activity (14). In addition, attacks must be accompanied by either nausea and/or vomiting or photophobia and phonophobia (14). Questionnaires such as the ID-Migraine (15) and 4-item migraine screener (16) have been developed and used as diagnostic screeners to aid in the diagnosis of migraine.

Medical claims databases, which generally store anonymized, individual-level, and standardized data on patients (e.g., age and sex) and claims (e.g., diagnosis, prescriptions, and treatments) in both inpatient and outpatient settings, are used in studies on many diseases in Japan; however, these data are not primarily generated for research purposes. Therefore, the use of validated algorithms to identify patients with a disease of interest is crucial to avoid misclassification of outcomes and exposures (17, 18), which can introduce bias (19). In validation studies, the accuracy of claims-based algorithms is usually compared to a gold-standard source of information such as clinical laboratory values, medical chart reviews, or registry data (20, 21). However, for migraine, which is diagnosed based on patients' descriptions of their symptoms, information obtained from a self-report questionnaire may help detect it. Indeed, a previous validation study conducted in the U.S. developed algorithms to identify patients who had not been diagnosed with chronic migraine using claims data. The study evaluated its performance using a questionnaire survey in combination with semi-structured interviews as the gold standard method for diagnosing chronic migraine (22).

In Japan, no studies have assessed claim-based algorithms for migraine detection. Therefore, in the present study, we compared 12 coding algorithms to identify people with migraines within a Japanese claim database based on four measures: sensitivity, specificity, positive predictive values (PPVs), and negative predictive values (NPVs). Given the diagnostic nature of migraine, which depends on patients' self-description of their symptoms, we used patient responses to an online survey asking about their symptoms, which incorporated questions in line with the diagnostic criteria of the ICHD-3. We considered migraine cases classified according to the survey responses as true migraine cases. Additionally, we have also examined the four measures considering individuals classified as having migraine according to the ID-Migraine and 4-item migraine screeners as true cases. Our findings could provide useful information to define and validate migraine cases in future database studies.

This study used a combination of administrative claims data and the linked results of an online survey. The combined data were obtained from DeSC Healthcare, Inc. (DeSC). The details of the conduct of this study were provided in an earlier publication (23).

The claims data, covering the period from 1 December 2017 to 30 November 2020, were provided by the society-managed employment-based health insurance associations that had contracts with DeSC for subscribers who agreed to the secondary use of their medical data by DeSC (~600,000 subscribers). Data included patient information (e.g., age and sex), diagnoses, prescriptions, and treatments in both inpatient and outpatient settings.

Regarding the survey data, an online survey was administered by DeSC to the registered users of the health app “kencom^®” aged 19–74 years (~150,000 users), irrespective of the presence of headache, from 1 to 30 November 2020. This health monitoring app was designed by DeSC and is freely available to users in Japan who are members of an affiliated society-managed, employment-based health insurance association (24). The survey questionnaire included items on sociodemographic characteristics (e.g., age, sex, residential area, occupation, and annual household income) and questions related to headache (the clinical features that include headache in the past 3 months, headache frequency, symptoms of headache, types of headache medicines used, impact of headache on daily activities, and questions to measure migraine-specific quality of life, work productivity, and activity impairment) (23). The questions also included items in line with the diagnostic criteria of the ICHD-3 (14) or common screening tools for migraine [e.g., ID-Migraine (15) and 4-item migraine screener (16)]. The study included all individuals who responded to the online survey regardless of whether they had headaches.

This study was approved by the independent ethics committee of Otsuka Pharmaceutical Co., Ltd. (approval no. 220617). The study used anonymized data, and no new individual-level consent was obtained for data use. The survey was conducted in accordance with the ethical guidelines for medical and biological research involving human subjects in Japan and the Declaration of Helsinki (revised in October 2013) of the World Medical Association.

In this study, we explored the ability of 12 coding algorithms, based on diagnostic records alone or in combination with prescription records, to identify people with migraine within a large-scale Japanese claims database. As migraine is diagnosed based primarily on patients' descriptions of their symptoms, we judged the “true” migraine cases using self-reported information collected in the online survey. The non-use of semi-structured interviews limited the accuracy of “true” migraine cases in this study. However, we observed consistent trends in the assessment of our algorithms for the three criteria used for case ascertainment. These trends, which are discussed below, can provide useful information for considering an appropriate coding algorithm to define migraines in each claims database study.

The 12 algorithms had overall low PPVs of ~20–30% with the ICHD-3 criteria and 30–40% with the ID-Migraine or 4-item migraine screener. However, this result should be interpreted in light of the fact that PPV depends on disease prevalence (26). These values were influenced by the low prevalence of migraine in this population (2.9–3.2%), which was lower than the reported prevalence of 6.0–8.6% in Japan (3, 4). This low prevalence may be related to the fact that this study was based on a self-administered survey without semi-structured interviews, which may have underestimated the prevalence of migraine and PPV. Many people with migraine are reportedly undiagnosed (3, 5, 27). A previous study reported that 59.4–71.8% of people with migraine in Japan had never consulted a physician previously, and only 11.6% were aware of the condition (28). Low awareness of migraine and the common use of OTC drugs for headaches (3) may have underestimated the number of people with migraine in the claims database. Indeed, only 250 people had a migraine diagnostic record in this study, whereas the responses to the survey suggested that 616–691 people possibly had migraines. This may also be partially responsible for the overall low sensitivity of our algorithms (5.4–13.5%), given the possibility of misclassifying patients who had migraines as “false negatives” because there are no migraine records in their claim records.

Among the 12 algorithms, we found that Algorithms 9 and 5 yielded the highest PPVs (32.5–42.1%), and this trend was consistent regardless of the criteria used for case ascertainment. This was probably because the strict conditions that require single (Algorithm 5) or multiple (Algorithm 9) diagnostic records plus prescription records of migraine prophylaxis or triptans could reduce “false positives.” Interestingly, the addition of acetaminophen or NSAIDs to triptans lowered the PPVs (Algorithms 6 vs. 5, 8 vs. 7, 10 vs. 9, and 12 vs. 11). These analgesics have wider indications and can be used for other conditions, which may have increased “false positives.” In contrast, Algorithm 2, which relied on a single diagnostic record of migraine or tension-type headache had the lowest PPV (19.2–27.1%). The algorithm used in a study that aims to estimate the relative risk of an outcome should have a high PPV so that the relative risk can be correctly estimated (26). Therefore, a strict coding algorithm based on a combination of multiple diagnostic records and prescription records of specific treatments, such as Algorithm 9, may be the most suitable for a migraine study with such an aim because it can reduce “false positives” and increase PPV.

However, a high PPV has high specificity, possibly at the sacrifice of sensitivity, as a result of increasing “false negatives (26).” Accordingly, Algorithm 9, which had high PPVs, had the lowest sensitivity (5.4–7.4%), while Algorithm 2, which had low PPVs, had the highest sensitivity (8.8–13.5%). For an exploratory study or a study that aims to estimate the prevalence or incidence of an outcome, an algorithm with high sensitivity is desirable (26) because maximizing the inclusion of patients with the target disease is more important than purifying the population under strict conditions. Therefore, for such a study aim, an algorithm based solely on a simple diagnostic record of migraine or tension-type headaches, such as Algorithm 2, may be desirable. This algorithm had the highest sensitivity, although at low levels, yet had an excellent specificity of ~99%, indicating the high accuracy of the diagnostic records in the claims database.

In stratified analyses, the PPV was higher in women and individuals aged 20 to 40 years, which is in line with the sex and age trends of migraine prevalence. The prevalence of migraine in women is 2 to 3.6 times higher than that in men (3, 29). Moreover, the prevalence is the highest between 30 and 39 years of age and decreases as age increases (30). Considering that the present study population was predominantly male (73.6%), the prevalence of migraine may have been lower than that in the general population, which may have affected the PPVs observed in our analysis of the overall study population. Therefore, our data stratified by sex and age group may be more useful as a reference when considering an appropriate claims-based algorithm in other studies, depending on the demographic characteristics of the population in these settings.

This study has some limitations, including those previously reported (23). For example, the study population may not represent the overall adult population of Japan since the database comprised data from employees and family members of large companies that are members of the health insurance associations, and survey respondents were limited to the Kencom^® users, with a high proportion of men (73.6%) compared to the general population in Japan [the prevalence of migraine in women and men was however 7.4 and 1.7%, respectively (23)]. These background characteristics may have influenced the low number of PPVs identified in this study. Because the prevalence of migraine or the distribution of other factors in our data may differ from those of other settings, the absolute values of this study may not be applicable to other databases.

One major limitation was that the present study used self-reported information collected in the online survey to judge the “true” migraine cases. Although the survey questionnaire included all items necessary to classify migraine according to the ICHD-3 criteria, no consultations with healthcare providers restricted us from obtaining an accurate or more reliable “true” status of migraine. Moreover, the ID-Migraine and 4-item migraine screener are screening tools although their use is recommended to aid diagnosis in clinical practice (25). However, these screening tools were used for supplemental purposes, considering that the ICHD-3 criteria may be too strict and miss unknown “true” migraineurs. Therefore, the accuracy of the case ascertainment in this study was limited, and the absolute values obtained should not be overinterpreted. Nevertheless, it is noteworthy that consistent trends were observed in the performance of the 12 algorithms across the three criteria. These findings will help us understand the advantages and disadvantages of each claims-based algorithm for use in migraine studies using Japanese claims databases.

A claims-based algorithm based on both the diagnostic records and prescription records of specific migraine medications had the highest PPV among the 12 algorithms considered, suggesting that such strict conditions may be appropriate for a study aimed at estimating relative risks. However, for an exploratory study, an algorithm based on a single diagnostic record of migraine or tension-type headache may be more suitable because it has higher sensitivity while maintaining high specificity. Because a suitable algorithm differs depending on the purpose of the study, it is important to choose an appropriate algorithm to define migraine in each claims database study.

The data that support the findings of this study are available from DeSC Healthcare, Inc. (Tokyo, Japan), but restrictions apply to the availability of these data, which were used under license for the current study and are not publicly available. However, data are available from the authors upon reasonable request and with permission of DeSC Healthcare, Inc. Requests to access these datasets should be directed to KY, Yamato.Kentaro@otsuka.jp.

The studies involving humans were approved by the Independent Ethics Committee of Otsuka Pharmaceutical Co., Ltd. (approval no. 220617). The study used anonymized data, and no new individual-level consent was obtained for data use. The survey was conducted in accordance with the Ethical Guidelines for Medical and Biological Research Involving Human Subjects in Japan and the Declaration of Helsinki (revised in October 2013) of the World Medical Association. The studies were conducted in accordance with the local legislation and institutional requirements. Written informed consent for participation was not required from the participants or the participants' legal guardians/next of kin in accordance with the national legislation and institutional requirements.

KY and HS contributed to the study conception, design, and contributed to the manuscript drafting. KH and TN reviewed and revised it critically for important intellectual content. All authors approved the final version of the manuscript for publication and agreed to be accountable for all aspects of this study and made substantial contributions to the analysis or interpretation of data.