Abstracts

Rationale: Accurate classification of electrographic seizures and identification of seizure onset times from intracranial electroencephalography (iEEG) recordings can facilitate physician iEEG review and provide information to support selection of treatment strategies that improve patient outcomes.

Methods: iEEG recordings from 191 patients with drug resistant focal epilepsy in the clinical trials of the RNS® System were analyzed, with each recording consisting of 4 channels. For electrographic seizure classification, data from 113 of these patients (~550,000 recording channels) were used. 72 patients were randomly chosen for training, 18 for validation and 23 for testing. For seizure onset time identification, data from 122 patients were used, with seizure onset times for ~8000 iEEG channels manually labeled. Data from 98 randomly chosen patients were used for training and the remaining 24 for testing. Due to the time-consuming nature of human labeling, the original dataset for seizure onset identification contained a limited number of training samples. To increase the size of the training dataset, two types of data augmentation were performed: 1) translational shifting of seizure onsets within individual iEEG channels by removing data from the end of the recording and replicating an equivalent amount of data at the beginning (~50,000 samples), and 2) translational shifting of seizure onset times by appending baseline non-seizure activity from randomly chosen iEEG files from the same patient to the beginning of the seizure recording (~215,000 samples). Various AI models were used, including convolutional neural networks (CNNs), vision transformers (ViTs), and time-series transformers. Vision-based models used time-frequency spectrogram representations as input, whereas time-series transformers processed raw time-series data. The models underwent comparison across different conditions, encompassing variations in model initializations (ImageNet vs random), datasets (original vs 2 augmentation types), and input types (colored vs grayscale spectrograms). Binary classification accuracy was used to compare the model performance for electrographic seizure classification. Median absolute error (MAE) between human labels and model predictions was used to compare performance for seizure onset time identification, since this metric is robust to diverse preferences of labelers.

Results: For electrographic seizure classification, ViTs emerged as the top performer, with a test classification accuracy of 96.9%, closely followed by CNNs achieving an accuracy of 96.2%. Both models used colored spectrograms as inputs. For the task of seizure onset time identification, CNNs showed a notably low MAE of 2.3 seconds, while for ViTs it was 2.6 seconds (performance with second augmentation type and ImageNet initialization).

Conclusions: With ~97% seizure classification accuracy and MAE of ~2.3 seconds in onset time identification, AI models may offer valuable tools to facilitate physician review of large iEEG data sets in order to develop personalized therapeutic interventions.

Funding: n/a