Abstracts

Rationale: Automated interictal epileptiform discharge (IED) detection can minimize the need for visual inspection, predict seizure recurrence when an increased frequency of the discharges is seen, [1] and identify the epileptogenic zone for surgical resection. Identifying interictal epileptiform activity is non-trivial since normal variants can be mistaken for IED with current machine learning techniques. Few-shot learning is a branch in deep learning that aims at efficient and automatic learning of relevant features from limited training data, while preventing overfitting. [2] This study proposes a novel deep learning methodology based on few-shot learning to automate IED detection with limited short episodes of IED samples from EEG data.

Methods: The proposed method is divided into 3 major steps: (i) First, using a symbolization technique named 1D-LBP, each channel in EEG data is converted into a feature vector which represents unique morphological patterns of non-linear EEG signal. (ii) Next, a deep neural network is trained to produce the pretrained model to classify IED vs. Non-IED events. Using weights of a pretrained network used to solve a similar task (i.e. seizure vs. non-seizure classification of CHB-MIT dataset), can improve the generalization capability of proposed system. (iii) Finally, using feature vectors from (i) and by fine tuning the pretrained network weights from (ii), few-shot learning is performed for binary classification of IED vs. non-IED events.

Results: 390 Scalp EEG recordings with annotations of six epileptic events (3 IED and 3 non-IED) were analyzed from TUH EEG dataset (https://isip.piconepress.com/projects/tuh_eeg/). Efficacy of proposed method was demonstrated by performing multi-fold cross validation for classifying IED vs. non-IED events. Duration of training data = 9744 seconds (each 1 sec. window contains one or multiple events). The proposed model achieved an average accuracy and specificity 81.0% and 81.7%, respectively. Preliminary results are promising and show the potential to detect IEDs from limited labeled data with high accuracy.  Currently, we are working to incorporate preprocessing and post-processing steps, as well as extending the architecture to multiclass classification.

Conclusions: Automated detection of IEDs can assist neurologists in diagnosis of epilepsy by providing temporal information related to seizure activity. Moreover, it can assist in analyzing the relationship between high occurrence of IEDs and seizure onset zones, which may benefit in planning surgical resection or implantation of neurostimulators. 

Funding: None