Non-invasive wearable seizure detection using long-short-term memory networks with transfer learning
Mona Nasseri1,2, Tal Pal Attia1, Boney Joseph1
1Bioelectronics Neurology and Engineering Laboratory, Department of Neurology, Mayo Clinic, Alfred 9-441C, 200 First Street SW, Rochester, MN 55905, United States of America.
Journal of Neural Engineering
|March 17, 2021
Summary
Wearable devices can now detect seizures using a deep learning algorithm trained on physiological data. This new method shows promise for improving epilepsy management, especially for motor seizures, with high accuracy in ambulatory settings.
Area of Science:
- Biomedical Engineering
- Neuroscience
- Artificial Intelligence
Background:
- Epilepsy management can be improved by detecting seizures with wearable devices.
- Reliable ambulatory seizure detection is challenging, and current methods lack electroencephalography (EEG) validation.
Purpose of the Study:
- To develop and validate a deep learning algorithm for seizure detection using wearable sensors.
- To assess the algorithm's performance with and without transfer learning from intracranial EEG (iEEG) data.
Main Methods:
- An adaptively trained long-short-term memory deep neural network was developed.
- Transfer learning adapted an iEEG-trained classifier for physiological datasets (accelerometry, blood volume pulse, etc.).
- Performance was evaluated using long-term ambulatory data with concurrent iEEG validation.
Main Results:
- The algorithm achieved high performance for motor seizures (AUC 0.98, sensitivity 0.93) in-hospital.
- Ambulatory detection of probable motor seizures showed good results (AUC 0.97, sensitivity 0.9).
- Detection of all seizure types in ambulatory settings yielded an AUC of 0.82, sensitivity of 0.47, and FAR of 7.2/day.
Conclusions:
- The algorithm can detect various motor and non-motor seizures.
- Performance is significantly better for motor seizures compared to non-motor seizures.
- This wearable-based seizure detection shows potential for epilepsy management.
