VILNIUS TECH Library invites you to follow the published new dissertations. The dissertation „Research and application of machine learning methods for migraine attack prediction“ prepared at VILNIUS TECH by Viroslava Kapustynska. The dissertation was prepared in 2021–2026. Scientific consultant – Prof. Dr Šarūnas Paulikas.
The dissertation was defended at the public meeting of the Dissertation Defense Council of the Scientific Field of Electrical and Electronic Engineering in the Aula Doctoralis Meeting Hall of Vilnius Gediminas Technical University at 2 p.m. on 9 June 2026.
Migraine is a complex neurological disorder characterized by strong inter- and intra-individual variability, which makes early forecasting difficult using only clinical observations. Wearable biosensors combined with machine learning offer new opportunities to detect subtle physiological changes that may precede migraine attacks and to develop individualized prediction models. This dissertation investigates migraine analysis and next-day prediction using physiological recordings collected under real-life monitoring conditions. Data were obtained with the Empatica Embrace Plus wearable device and include electrodermal activity, pulse rate, skin temperature, and movement-related signals. The analysis focuses on nocturnal recordings, since the night period provides a more stable physiological context with fewer external disturbances. Nights were standardized using sleep-based contextual selection and consistent night-level rules. The experimental framework is organized in two stages. In the first stage, a window-level binary classification task is used as an exploratory methodological analysis to examine how design choices influence model performance. Night recordings are segmented into analysis frames ranging from 5 to 120 minutes, statistical features are extracted, and the influence of signal preprocessing and feature representation is evaluated across several classifier families, including Random Forest, XGBoost, histogram-based gradient boosting, support vector machines, and k-nearest neighbors. In the second stage, the research evaluates next-day migraine prediction based on whole-night recordings. This stage refines the experimental methodology to obtain more reliable estimates of predictive performance under a stricter validation framework. The analysis focuses on the effect of temporal aggregation while comparing the same classifier families under consistent evaluation conditions. The results demonstrate considerable variability across participants in achievable prediction performance and optimal modeling configurations. Shorter analysis frames generally preserve informative short-term physiological changes, whereas longer windows tend to smooth these variations. Signal preprocessing shows a window-dependent effect and does not consistently improve performance. Overall, the results highlight the importance of temporal resolution, rigorous validation, and individualized modeling for wearable-based migraine prediction systems.
Doctoral dissertation readers can search via VILNIUS TECH Virtual Library.
