Predicting Migraine Onset Using Machine Learning and Daily Lifestyle Data

Migraines are a widespread neurological condition with significant personal and societal impact, yet their unpredictability remains a challenge. This thesis investigates whether migraine attacks can be predicted one day in advance using machine learning techniques. The study is conducted in collaboration with Neurawave and utilizes two datasets: the first consists of daily health reports from four individuals overapproximately 90 days; the second comprises daily diary entries from over 60 participants in a Korean migraine study. To address the limitations of the first dataset, additional features such as weather conditions, moon phases, and air pollution are integrated. Temporal patterns are captured through lag features, and Recursive FeatureElimination with Cross-Validation (RFECV) is applied to identify relevant predictors. A range of supervised models, including Support Vector Machines, RandomForest, XGBoost, Logistic Regression, and a Neural Network, are trained and evaluated using AUC-ROC and Brier scores. While the best-performing model on the first dataset achieves an AUC-ROC of approximately 0.69, the results are not sufficient for clinical or commercial use. However, experiments on the alternative dataset show significantly improved results for same-day prediction, with all tested models reaching AUC-ROC values above 0.80 and Brier scores below 0.16. These findings suggest that short-term migraine prediction may be feasible with current data and models, and that improved data collection could further enhance predictive accuracy. This work provides a foundation for future research in personalized migraine prediction systems.