Authors :
Presenting Author: Mona Nasseri, PhD – University of Northern Florida
Rachel Stirling, ME – University of Melbourne; Pedro Viana, MBBS – King's College London; Jie (Richard) Cui, PhD – Mayo Clinic; Ewan Nurse, PhD – Seer Medical; Matthias Dümpelmann, PhD – Freiburg University; Dean Freestone, PhD – Seer Medical; Mark Richardson, PhD, FRCP – King’s College London; Benjamin Brinkmann, PhD – Mayo Clinic
Rationale: The unpredictability of seizures can be debilitating and dangerous for people with epilepsy. Accurate seizure forecasters could improve quality of life for people with epilepsy but imust be practical for long-term use. Here we demonstrate pseudo-prospective forecasts with non-invasive peripheral wearable devices with seizure confirmation from concurrent chronic EEG recordings.
Methods: Patients were recruited for ultra-long term monitoring with a wearable device (Empatica E4/ Fitbit Inspire HR or Fitbit Charge 3) and concurrent chronic EEG monitoring (UNEEG SubQ, NeuroPace RNS or Medtronic Summit RC + S) at Mayo Clinic, Rochester MN and King’s College London. Physiological data was recorded from enrolled patients for at least six months. EEG data was reviewed and seizures were confirmed by neurophysiologists. Wearable devices recorded various signals, however only step counts and heart rate signals were used to train two forecasting algorithms; short-term memory (LSTM) recurrent neural network and a cycles-based models. Final forecasts were an ensemble of those classifiers, which combined short (minutes to hours) and long (days to months) horizons.
Results: Ten participants with epilepsy were included in the forecasting analysis. An average of 311 days of wearable monitoring and 65 seizures were recorded per patient. LSTM forecasts performed better than chance (p < 0.05) for eight of ten participants, with an average AUC score of 0.67 and AUC of 0.63 across all subjects. Cycles-based forecasts performed better than chance for all participants, with an average AUC score of 0.68 across all subjects. The ensembled forecast model improved forecast scores for three participants.
Conclusions: This study demonstrated the feasibility of forecasting electrographic seizures with wearable devices and the possibility to enhance the performance using an ensemble model which is the combination of two established forecasting techniques: LSTM models for short horizon (minutes to hours) forecasting and cycles-based models for long horizon (days to weeks) forecasting.
Funding: Epilepsy Foundation of America My Seizure Gauge