LAUSR

In a highly innovative model development study, Pavlovic and colleagues (1) developed a claims-based prediction model to identify individuals with undiagnosed chronic migraine. Participants in the study were adults recruited from a large medical group and were only eligible for inclusion if they possessed a diagnostic code for migraine (excluding chronic migraine) and were being followed by either the medical or pharmacy services in the year preceding the database query. Out of 40 predictor candidates, four predictors of undiagnosed chronic migraine were isolated that included number of claims for acute treatment of migraine (including opioids), number of healthcare visits, being female, and claims for unique migraine preventive classes. When combined, these four variables possessed extremely promising predictive power as they discriminated between those with and without chronic migraine (area under curve (AUC): 0.80). In addition to being highly innovative, this study exhibited several particular strengths. The authors successfully labored to identify cases of true chronic migraine from the pool of participants. To accomplish this, they employed an additional screening tool (ID-CM (2)) combined with a semi-structured diagnostic interview. Using these methods, a reliable diagnosis of chronic migraine could be assigned to eligible participants. With the growing popularity of machine learning approaches for medical classification problems (3), there has been an increase in prediction models using data sets like that available to Pavlovic and colleagues. However, the vast majority of these efforts will fail to have such an exemplary gold standard as the one utilized in this study. In claims databases, diagnostic codes or procedural codes cannot often be assumed to reflect the construct of interest (e.g. a refined diagnosis). Thus, obtaining a high level of reliability in the outcome requires great forethought and, at times, actual records review or additional interviews. The extra effort taken to assure accurate diagnosis is a major strength of this study. A second major strength of this study is the blending of theory-based model selection with data-driven approaches. As the authors admit, their available sample size was modest for such an effort, but by blending a bivariate screening approach with an expert consideration of potential predictive validity, a parsimonious set of predictors could be identified without substantial correlation among the predictors (i.e. collinearity). This same approach could be reapplied with the advent of additional data to increase the number of predictors, increasing the predictive power of the model even further. Each of the predictors in the final model exhibits substantial face validity, and each exhibits an independent association with undiagnosed chronic migraine. A major problem with prediction models is that many of them do not replicate in future studies. By using both data and theory to develop their model, Pavlovic and colleagues substantially increase the chances that their model will be useful in other claims databases. Despite these impressive strengths, the most exciting aspect of this study is simply the vast potential of the prediction model to reduce undiagnosed chronic migraine. The specified predictors are easily measured within any claims database, and the predictors in the model were dichotomized to ensure that the model could be deployed without a sophisticated algorithm. Thus, with the initial validation study completed, the model can be readily employed to predict the probability that individuals receiving treatment for migraine are actually experiencing chronic migraine. Epidemiological studies have shown that only 20 to 25% of individuals with chronic migraine actually receive an accurate diagnosis (4,5). This prediction model offers a rare opportunity to identify at-risk individuals who are already seeking treatment for refinement of diagnosis and improved treatment.