Advancing a generalizable model for migraine prediction: Analysis of filtering techniques on physiological signals

Despite wearable sensors' ability to provide continuous physiologic monitoring, migraine remains challenging to predict due to unpredictability of onset and a variety of triggers. Developing an accurate prediction model requires reducing signal variability by using effective filtering techniques.

The main objective of this study is to evaluate machine learning models for predicting migraines and analyze the effect of different filtering techniques and classifiers on prediction performance.

A feature set based on ANOVA analysis of four key physiological signals was used. After the pre-processing, filtering methods, including median, Butterworth, and Savitzky-Golay filter, were applied. Five classification models, Extreme Gradient Boosting, Histogram-Based Gradient Boosting, Random Forest, Support Vector Machine, and K-Nearest Neighbors, were evaluated.

The highest predictive performance was achieved using the Savitzky-Golay filter. The Random Forest model demonstrated the best accuracy (0.858) and precision (0.815), and an F1-score of 0.677, indicating the potential of investigated signals for migraine prediction. Furthermore, the Histogram-Based Gradient Boosting model achieved the highest recall using the Savitzky-Golay filter (0.719), demonstrating its effectiveness in identifying true positive cases of migraines.

The results indicate significant potential for healthcare applications for early migraine prediction and treatment using wearable technology. The study highlights the importance of selecting appropriate features and filtering methods to improve the accuracy and reliability of the predictions.