Abstracts

Rationale: Seizure forecasting could significantly improve the quality of life of patients with epilepsy (PWE). While earlier investigations have focused on intracranial electroencephalography recordings, wearables could offer more practical, affordable, and accessible solutions. Sleep is an interesting biomarker given its bidirectional relationship with seizures and the ability to measure it non-invasively [1]. The aim of this study was to investigate if seizure days can be forecasted based on nocturnal sleep recordings of a smart shirt [2].


Methods: Seventy-eight PWE admitted to the University of Montreal Hospital Center Epilepsy Monitoring Unit (EMU) wore the Hexoskin biometric smart shirt during their stay. Using continuous cardiac, respiratory, and movement recordings, the following metrics were calculated by the smart shirt built-in sleep algorithm [3] for each night (total of 422 nights): sleep efficiency, sleep latency, total sleep duration, time in rapid eye movement (REM) and non-REM sleep, wakefulness after sleep onset, average heart rate and breathing rate during sleep, average high-frequency heart rate variability, and the number of position changes during the night. The ten features were used as input to a supervised machine learning algorithm which classified each day (24-hour period starting at wake) as a seizure day or non-seizure day. Seizure days were determined by neurologists using video-EEG recordings as 24-hour periods with at least one clinical seizure. After normalizing the data for each PWE with an individualized reference night, a binary support vector machine classifier was trained for each of the 45 PWE who had seizure days, using a nested leave-one-subject out cross-validation approach. Hyperparameter tuning was performed in the inner cross-validation loop.


Results: Improvement over chance (IoC) seizure-day forecasting performance, defined as the difference between the sensitivity and the percentage of time in warning (TiW) [4], was achieved in 40.0% of PWE (18 of 45 PWE). When considering only PWE with IoC performance, the mean IoC, sensitivity, and TiW were 34.2%, 64.4% and 30.2%, respectively. Our results are comparable to a recent EMU study, where IoC forecasting was achieved in 43.5% of PWE (30 of 69) using a wristband sensor [4].


Conclusions: Our results show promise for noninvasive seizure-day forecasting in PWE at the EMU using smart wear-based nocturnal sleep monitoring. Further studies in a residential setting with long-term recordings could lead to the development of novel and practical seizure advisory devices.


Funding: This work was supported by IVADO, iTMT, CIHR, NSERC, FRQNT, and the Canada Research Chair Program (DKN).

References


[1] Frauscher, B. & Gotman, J. Neurobiol Dis. 2019;127: 545-553.

[2] Ding, TY. et al., Epilepsia Open 2024. Under review.

[3] Pion-Massicotte, J., et al. J Sleep Res. 2019;28(2), e12667.

[4] Meisel, C. et al., Epilepsia. 2020 Dec;61(12):2653-2666.