The FORESITE Study: FOREcasting Seizures to Initiate ThErapy – Technical Design, Hardware, and Algorithms
Benjamin Brinkmann1, Jie Cui2, Jordan clark1, Mona Nasseri3, Pedro viana4, Vaclav Kremen1, Jonas Duun-Henricksen5, Kevin McQuown6, Rachel Stirling7, Vlad Sladky1, Phillipa Karoly7, Enrique Carrazana8, Andreas Schulze-Bonhage9, Mark Richardson4, Gregory Worrell2
1Mayo Clinic, 2Mayo Clinic College of Medicine, 3Electrical Engineering, University of North Florida, 4Institute of Psychiatry, 5UNNEG Medical, 6Windy City Labs, 7Biomedical Engineering, University of Melbourne, 8Neurelis, Inc., 9Universitaetsklinikum Freiburg
Objective:
For seizure forecasts to improve the lives of people with epilepsy, forecasts may be used to prompt medication administration to reduce seizure burden with minimal increase in medication dose.
Background:
Seizure forecasting has been demonstrated with intracranial EEG, wearable devices, and subscalp EEG. Our team is developing a seizure forecasting system based on subscalp EEG and wearable device signals operating via a smartphone and an associated microcontroller. Seizure alerts prompt administration of intranasal diazepam, with the goal of preventing seizures before they begin. Data is processed on an independent microcontroller (Particle Photon2) connected via Bluetooth to a subscalp EEG device (UNEEG SubQ, Allerod DK). Heart rate and actigraphy are recorded from a smartwatch. Forecasting algorithms run on the microcontroller, and recorded data is transmitted to the cloud. Push notifications are sent to the patient’s phone in response to seizure forecasts to prompt medication administration. Doses are logged on a smartphone seizure diary, and further alerts are suppressed once the dosage limit is reached. Detected seizures on EEG prompt cognitive testing on the smartwatch to test impaired consciousness.
Design/Methods:
Seizure forecasting algorithms use deep-learning LSTM neural network and multi-day cycle models of seizure risk. Given the benzodiazepine’s half-life (49.2h), circadian seizure risk patterns are not helpful, and multi-day cycles are necessary. Seizure forecasting algorithms were tested on 7 ultra-long-term SubQ EEG recordings from King’s College London. Subcutaneous EEG, annotations, and wearable signals are displayed in a cloud data dashboard.
Results:
A 2-class CNN-biLSTM seizure detection algorithm trained on a combined dataset of subcutaneous EEG and scalp EEG data produced an AUROC of 0.959 and AUPRC of 0.485 with positive sample fraction of 0.00021. Multi-day cycle seizure forecasting models achieve over 80% sensitivity with false alarms occurring on fewer than half of days in four of 7 patients.
Conclusions:
Seizure forecasting to guide medication is feasible.
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.