Unsupervised Detection of Migraine-Related Anomalies in Nocturnal Blood Volume Pulse Dynamics Using a Convolutional Autoencoder
Rūta Jankevičiūtė, Vytautas Abromavičius, Saulius Andriuškevičius
DOI: http://dx.doi.org/10.15439/2026F8288
Citation: Rūta Jankevičiūtė, Vytautas Abromavičius, Saulius Andriuškevičius (2026). Unsupervised Detection of Migraine-Related Anomalies in Nocturnal Blood Volume Pulse Dynamics Using a Convolutional Autoencoder. In M. Bolanowski, M. Ganzha, M. Grzegorowski, L. Maciaszek, M. Paprzycki, A. Paszkiewicz, D. Ślęzak (eds), Proceedings of the 21st Conference on Computer Science and Intelligence Systems (FedCSIS). ACSIS, Vol. 47, pages 537–542.
Abstract. This study investigates whether nocturnal blood volume pulse (BVP) dynamics exhibit anomalous patterns during migraine days. Rather than framing migraine detection as a supervised classification problem, we adopt an unsupervised anomaly detection approach to identify deviations from patient-specific baseline sleep physiology. Participants were adults (18+ years) who met the criteria of the International Classification of Headache Disorders (ICHD-3) for episodic migraine and reported between 4 and 14 migraine days a month. For each participant, a longitudinal BVP recording containing multiple nights of sleep data was analyzed in conjunction with a migraine diary. Participants with migraine data (n = 17) were compared using aggregated anomaly metrics. For most participants, the p99 ratio was higher with the 30 s / 5 s configuration compared to the 60 s/10 s setup, indicating greater separation between migraine and non-migraine days at finer temporal resolution. The results indicate that migraine days are generally associated with increased anomaly scores in sleep-time BVP signals. In addition, the majority of the participants exhibited a greater sensitivity to migraine-related changes in BVP dynamics. Overall, these findings support the use of a convolutional autoencoder for detecting migraine-associated deviations in sleep time BVP signal dynamics, particularly in settings where labeled data are limited and inter-subject variability is high.