OBJECTIVES : Chronic migraine (CM) and episodic migraine (EM) are part of the spectrum of migraine disorders, but they are distinct clinical entities. EM is characterized as up to 14 migraine days per months, while CM is defined as 15 or more headache days per months, at least 8 out of which have to be typical migraine headache days. We seek to use a machine learning technique to produce the classification hence infer the severity of the disease.

METHODS : The diagnosis codes in the MarketScan database were used to produce three cohorts containing CM, EM and non-headache patients (a negative control group). Features were derived and a list of supervised machine learning algorithms were explored including random forest, neural networks, gradient boost, and stacked model. Particle swarm optimization algorithm was applied to explore the optimal parameters of the machine learning algorithms.

RESULTS : 395 features were constructed. Each CM, EM or negative control group had 18,832 patients. The gradient boost algorithm produced the best result, which reported the accuracy and kappa at 0.75 and 0.63, respectively. After applying thresholds in the optimized gradient boost model, the accuracy and kappa were further increased to 0.81 and 0.72 with prediction decision made to 80% of the dataset. Collinearity and high correlation were found with features within the group regarding cost, resource, and medication use. 27% of CM and 9% of EM patients received wrong predictions from the fine-tuned random forest, gradient boost, and neural network. Migraine-related costs were ranked as top important features.

CONCLUSIONS : Migraine patients with higher cost, resource use and medications use were more likely to have a diagnosis of CM vs EM, with higher migraine-related costs as the top indicators of a patient having CM. Further research on a data source with more migraine specific clinical outcomes is deemed.