2024 Harold Wolff-John Graham Award: An Award for Headache/Facial Pain Research
Abstract
Background: The ability to forecast migraine attacks could offer patients opportunities for early intervention and activity planning. The KDT is a timed rapid number naming task that requires complex cerebral functions. Our goal is to construct a machine learning model that forecasts impending or continuation of migraine attacks based on detailed symptoms and King-Devick test (KDT) scores.
Methods: This prospective observational study evaluated 30 adults with migraine with or without aura who completed at least three diary entries and KDT daily for up to 4 months. Diary entries included a self-report of current migraine phase and detailed symptoms experienced. KDT score was time in seconds to complete the test with zero errors. We trained a machine learning model to forecast having a migraine attack during the next recording (in approximately 6–12 h) using 272 features including clinical symptoms and KDT scores. Shapley Additive exPlanations (SHAP) analysis was employed to determine the top 20 most important features in the model development.
Results: A total of 4970 data points containing detailed symptoms and KDT scores from 30 participants (mean age 44, 29 females) were recorded, including 3694, 518, 288, 181, and 289 during migraine attacks, non-migraine headache, prodrome, post-drome and interictal phase, respectively. Our random forest model achieved an AUC of 0.70. SHAP analysis suggests that during migraine attacks, fatigue, photophobia, neck stiffness, difficulty with concentration and cognition are important for attack continuation. Reporting photophobia, phonophobia and fatigue without headache during prodrome are suggestive of having impending attacks. For all migraine phases, shorter time from the last attack, longer KDT score, and shorter duration of sleep the night before are suggestive of having a migraine attack at the next recording.
Conclusion: Our effective machine learning model suggests photophobia, fatigue, lack of sleep and the KDT score, an objective tool, are the top features to forecast impending, or continuation of migraine attacks.