Abstracts

Rationale:
Reliable methods to assess seizure risk could alleviate a major burden for people with epilepsy(PWE) byprovidingtimelyseizurerisk warning.While recent work demonstrated that seizure risk assessment is possible,priorapproachesreliedlargelyoncomplex,ofteninvasivesetupsincluding intracranial electrocorticography, implanted devices and multi-channel EEG, and required patient-specific adaptation or learning to perform optimally. To facilitate broader adaptation, non-invasive,easily applicable techniques that reliably assess seizure risk without much prior tuning arecrucial. Such techniquesaredesirablebecauseoftheirabilitytoimprovetreatmentbyoptimizingdosingandtimingofantiseizure medicationregimenutilizingpersonalizedstandards,andpotentiallyenablingtimelyinterventionstoavertimpendingseizures, minimizing potential complications associatedwithinvasivemonitoring,andavoidingstigmaassociatedwithbulkyexternal monitoring devices on thehead.
Method:
We applied deep learning (LSTM and 1DConv networks) on multi-modal wristband sensor data from 69 PWE (mean age 10±6 years, 28 female, total duration 2311.4 hours, 452 seizures) to assess its capability to forecast seizures in a clinically meaningful way. PWE were equipped with a wireless multi-sensor device (Empatica® E4, Milan, Italy), which records temperature, photoplethysmography, electrodermal activity and actigraphy, during long-term, continuous video-EEG monitoring. We applied a leave-one-subject-out cross-validation approach where matched pre-/interictal data (classified based on 30-second data segments) were used for training, and testing was done on the remaining out-of-sample patient dataset. We evaluated performance of the proposed seizure forecasting system in terms of sensitivity, time in warning and improvement over chance (IoC). Results Wefound seizure forecasting to be significantly better than chance for 43.5% of patients (30/69 patients) yielding a mean IoC of 28.5±2.6% and a mean sensitivity of 75.6±3.8% (Fig. 1). The mean prediction horizon was 1896±101 seconds, a period that may be long enough to afford reasonable warning in advance. Control analyses based on time-matched seizure surrogate data indicated that forecasting was not simply based on time of day or vigilance state. To better understand each sensor modality’s contribution to successful seizure forecasting, we also performed analyses by removing each sensor’s data individually, which indicated that all data streams contributed to seizure forecasting.We found that prediction performance increased with size of the training dataset indicating potential for further improvement with larger datasets.
Conclusion:
Results demonstrate the capability of multi-modal wristband sensor data from easy-to-use, non-invasive devices in combination with deep learning to provide statistically significant and clinically useful seizure forecasting. These initial results may provide the basis for future re-evaluation, algorithm improvement and benchmarking as a step towards patient empowerment and objective epilepsy diagnostics for broad application.
Funding:
:Epilepsy Research Fund, NARSAD Young Investigator Grant