Abstracts

Rationale: Non-invasive wearable data may provide more precise seizure frequency assessments for improved management of patients with epilepsy. Current clinical practice primarily relies on patient self-reports, which may miss seizures. Automated detection of seizures from wristworn signals may improve seizure detection but to date mostly rely on short periods in the range of seconds. We aimed to evaluate seizure detection based on deep learning algorithms from prolonged raw data segments in an attempt to leverage longitudinal data signatures. _x000D_
Methods: We cut long recording in 60 minute segments of multi-modal signals collected with a wearable device (Empatica E4) during continuous video-EEG monitoring in the EMU. We used deep learning (conv-1D) to detect tonic-clonic seizures and investigated useful data signatures by visualizing the input data based on its contribution to seizure detection. First, we quantified potential artifact contamination and excluded periods of low-quality data from further analysis using a validated algorithm. Next, we fit models on full-hour data using a leave-one-patient-out 10-fold nested cross-validation approach. Finally, the best models were applied to the entire data from the left-out patients to predict the labels by a majority vote. To gain insight into what data features drove the detection algorithm, we used UMAP (Uniform Manifold Approximation and Projection for Dimension Reduction) to visualize feature separation and shapley values to visualize contributions of each data modality._x000D_
Results: We included multi-day recordings from 20 patients with epilepsy (14±4 years, 10 females, 103 seizures, total duration of 1721 hours). Tonic-clonic seizures were detected significantly better than chance (Figure 1 shows an example from one patient). Figure 2 demonstrates the separability of the learned features after applying dimensionality reduction with UMAP. Feature importance derived from the deep learning model’s shapley values pointed to a distinct physiological profile of tonic-clonic seizures: (1) rapid movements indicative of the clonic phase detected by actigraphy, (2) a post-ictal peak in electrodermal activity and (3) a peri-ictal low-to-high-increase in heart rate. _x000D_
Conclusions: Longer periods contain more information than has been used in previous seizure detection approaches. These long-term trends, including increases in heart rate and EDA peaks, provide purely data-driven insight into seizure physiology and help to train classifiers using the most informative signals._x000D_
Funding: CM was supported by a NARSAD Young Investigator Grant. This study was supported by the Epilepsy Research Fund. SV was supported by Deutsche Forschungsgemeinschaft, Grant/Award Number: VI 1088/1-1.