Abstracts

This abstract is a recipient of the Jack M. Pellock Pediatric Travel Award

Rationale: Proof-of-principle studies suggest that seizure prediction from non-invasive device recordings is feasible. However, the discovery of biomarkers is an ongoing research task. Our biomarker discovery work is based on the central hypothesis that impending seizures alter the central control of the autonomic network, causing coupled changes in autonomic subsystems. We aim to evaluate seizure-related changes in peripheral neurophysiological recordings.

Methods: We included 21 patients with generalized tonic-clonic (GTCS) or focal to bilateral tonic-clonic (FBTCS) seizures and 21 age and sex-matched controls that were admitted for continuous video-EEG monitoring at Boston Children’s Hospital and wore a wearable biosensor (Empatica^®, Milan, Italy) recording heart rate (HR) and electrodermal activity (EDA). EDA was z-transformed and 45 minutes of pre-ictal data and 45 minutes of time-matched control data from each patient’s matching partner were used for analysis. Those 45-minute-episodes were cut into 200 seconds segments with no overlap (leaving a 100-second buffer before seizures) and analyzed by deep canonically correlated autoencoders. More specifically, each modality’s data was fed into a deep neural network which maximizes the correlation between the lower-dimensional representations (outputs) from both neural networks in an unsupervised fashion. These maximally correlated representations were clustered into two different groups using spectral clustering. The clustering accuracy for matching the class labels and cluster labels was reported (Figure 1A, methods overview).

Results: Mean levels of EDA did not differ between groups (t(40) = 0.478, p = 0.637), while HR was lower (t(40) = -3.123, p = 0.003) for patients with seizures in pre-ictal compared to patients without seizures in inter-ictal time-matched segments (Figure 1B, group raw data). Clustering accuracy into interictal and pre-ictal groups was above chance level for 64% of the patients and above 0.9 for 43% of patients (Figure 1C).

Conclusions: Seizure prediction utilizing unsupervised clustering based on deep canonical correlation analysis of coupled changes in heart rate and electrodermal activity is feasible in a majority of patients with GTCS and FBTCS. Next steps include further validation including larger patient series, refining neural networks and the clustering methods, serial neurophysiological signal analysis, and the inclusion of additional variables.

Funding: This study was supported by the Epilepsy Research Fund. SV was supported by Deutsche Forschungsgemeinschaft, Grant/Award Number: VI 1088/1-1.