Abstract
Analysis of EEG requires years of clinical training and mentorship. To alleviate the human cost involved in EEG analysis, we propose a general LSTM-Autoencoder-CNN for EEG (GLACE) framework, which is adequately general to facilitate the use of transfer learning in smart healthcare. Traditionally in transfer learning, only the last few layers of the neural network are changed and adapted to the new task. Instead, we focus on the adaptation of the first layers to each new task. We exploit the inter-trial couplings in our proposed deep learning approach called GLACE. The efficacy of GLACE was assessed against a real-world clinical problem, i.e. the detection of interictal epileptiform discharges; GLACE circumvents the need for the neurophysiologist to spend hours on EEG analysis.
Simulations show that the adaptation of the first layers of the trained model leads to an accuracy improvement of 12%.
Simulations show that the adaptation of the first layers of the trained model leads to an accuracy improvement of 12%.
Original language | English |
---|---|
Title of host publication | International Conference on e-Health and Bioengineering |
Publisher | IEEE |
ISBN (Electronic) | 978-1-6654-4000-4 |
ISBN (Print) | 978-1-6654-4001-1 |
DOIs | |
Publication status | Published - 19 Nov 2021 |