Abstracts

Rationale: Recent advancements in artificial intelligence have led to the application of deep learning to epileptic seizure detection. While these methods have the potential to aid clinicians in finding seizure intervals in long-term monitoring settings, they fail to aid in the difficult task of seizure onset zone (SOZ) localization. Here we present GraphTrack, an end-to-end neural network for detecting and localizing seizure activity in clinical scalp electroencephalography (EEG) recordings. Using a novel combination of biologically inspired convolutional, graph, and recurrent neural networks, GraphTrack is capable of identifying and tracking the propagation of seizure activity at the resolution of one-second windows and individual EEG channels. Our approach opens a new direction for machine learning methods in epilepsy, by going beyond the standard seizure detection paradigm into the difficult but clinically relevant task of SOZ localization.

Methods: GraphTrack is built with a biologically inspired combination of Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Graph Neural Networks (GNNs). First, a three-layer CNN is used to encode a sequence of latent representations based on one-second windows of each EEG electrode channel. Second, the sequences of latent encodings are analyzed using an RNN, which captures temporal dependencies in the data and effectively "tracks" the evolution of seizure activity. Third, a GNN compares the representations in each channel to its neighboring and contralateral channels to identify groups of electrodes exhibiting correlated seizure activity. Finally, electrode-level seizure predictions are made for each one-second window of EEG, allowing GraphTrack to detect seizures at high temporal resolutions as well as predict SOZ.

Results: GraphTrack is evaluated on a dataset of 34 patients with focal epilepsy recorded at the Johns Hopkins Hospital. Training is performed using leave-one-patient-out cross validation, ensuring that GraphTrack is capable of performing well on unseen patients. In a seizure detection task, GraphTrack demonstrated a sensitivity of 0.88, which is comparable to state-of-the-art seizure detectors. In a localization task, GraphTrack correctly lateralized the SOZ with an accuracy of 0.85 and identified anterior (frontal lobe) versus posterior (temporal and parietal lobes) onsets with an accuracy of 0.82. The concurrence between the GraphTrack predictions and clinician annotations demonstrate its ability to track evolving seizure activity in focal epilepsy.

Conclusions: GraphTrack introduces a new task paradigm for joint seizure detection and SOZ localization in scalp EEG recordings. Early results show efficacy in detection, tracking, and localization, combining three clinical goals in epilepsy monitoring in one automated system. GraphTrack's unique approach to seizure activity tracking demonstrates its potential to become an exciting new tool for long-term EEG monitoring.

Funding: Please list any funding that was received in support of this abstract.: This work is supported by the National Science Foundation CRCNS award 1822575 and CAREER award 1845430, and the Johns Hopkins University Discovery Award.