Automated Assessment of Sleep in Patients with Dravet Syndrome from Simulated Behind-the-ear EEG
Abstract number :
1.53
Submission category :
3. Neurophysiology / 3C. Other Clinical EEG
Year :
2024
Submission ID :
1601
Source :
www.aesnet.org
Presentation date :
12/7/2024 12:00:00 AM
Published date :
Authors :
Benjamin Wittevrongel, PhD – Epilog, Clouds of Care
Tanja Hellier, PhD – UCB
Jan Vandenneucker, MSc – UCB
Julie Nys, PhD – Byteflies
Anouk Van de Vel, MD – University Hospital Antwerp
Riëm El Tahry, MD, PhD – UC Louvain
Sarah Weckhuysen, MD, PhD – University Hospital Antwerp
Caroline Neuray, MD – Epilog, Clouds of Care
Presenting Author: Pieter van Mierlo, PhD – Epilog, Clouds of Care
Rationale: Poor sleep is an under-evaluated but highly relevant comorbidity in children with epilepsy. A typical montage to assess sleep in a clinical setting uses an elaborate setup of EEG, ECG, EMG and EOG electrodes as well as respiration measurement. This is not only intrusive, especially for children, but also limits examinations to single-night sessions. To facilitate long-term monitoring, including daytime naps, a less intrusive approach is required, such as a behind-the-ear electrode setup that is well tolerated by subjects and can be discretely used during the day. This study explores the feasibility of automated pediatric sleep staging from simulated behind-the-ear EEG.
Methods: We first determined how to simulate behind-the-ear EEG from full-scalp EEG. We used 9 co-recordings containing both behind-the-ear and scalp EEG and obtained the averaged ordinary least-squares solution across the recordings to optimally match the recorded behind-the-ear EEG. Next, we collected a train set comprising fifty 24-hour epilepsy-free EEG recordings from Saint-Luc University Hospital (Brussels, Belgium), manually sleep-scored (Wake, N1, N2, N3, REM) by 3 independent neurophysiology experts. The test set consisted of twenty 24-hour EEG recordings from patients with Dravet syndrome (DS) from University Hospital Antwerp. Recordings were manually sleep-scored by 1 expert. For each of the 2 montages (i.e., full-scalp and behind-the-ear), we trained a sleep staging model on the train set using expert consensus scoring. Model performance was assessed by Cohen’s kappa (k) between the model's classifications and the scoring of the 1 expert. For performance on the control cohort, we used 5-fold cross-validation, and performance for the DS cohort was extracted directly. Differences between the modalities were assessed using (2-sided) Wilcoxon rank-sum test, and differences across modalities using (2-sided) Wilcoxon signed-rank test.
Results: The full-scalp model demonstrated good agreement with the expert scoring for both the control (κ=0.82±0.09) and DS (κ=0.77±0.21) cohorts. The difference was statistically significant (P=0.022). The behind-the-ear model also reached considerable agreement on the control (κ=0.76±0.10) and DS (κ=0.69±0.18) cohorts (P=0.038). The difference between the full-scalp and behind-the-ear models was significant for the control group (P< 0.001) but not for the DS cohort (P=0.277).
Conclusions: This study demonstrates that sleep can be reliably assessed from a behind-the-ear EEG montage in an automated way in both healthy subjects and patients with DS. Compared to the full-scalp solution, the behind-the-ear montage results in a slightly decreased performance, but the hypnograms maintain substantial agreement with expert scoring. The slight decrease in performance in the DS cohort can be explained by the fact that the neural signals are more contaminated due to epileptic abnormalities. Nevertheless, this study has shown the potential for long-term sleep monitoring using non-intrusive wearable hardware.
Funding: UCB Pharma funded this work.
Neurophysiology