Abstracts

Rationale: The rise of Machine Learning (ML) has prompted a shift in epilepsy research towards spatio-temporal predictive analyses. Common applications of ML include automated seizure detection, seizure type classification, and seizure onset localization. While ML algorithms are growing in popularity, they are often “black boxes” that lack interpretability. One way to address this limitation is to visually compare the EEG data with the model predictions. In the case of epilepsy, these visual comparisons can help determine whether the algorithm is responding to clinically meaningful features or artifact. Current EEG software packages do not provide the required visualization capabilities.

Methods: We introduce a novel open-source viewer, the EEG Prediction Visualizer (EPViz), to aid researchers in developing, validating, and reporting predictive model outputs. EPViz is a standalone software package developed in Python and targeted towards ML applications. It is built around four core functionalities: (1) displaying multi-channel EEG signals, (2) running PyTorch deep learning models on EEG, (3) overlaying channel-wise and time-varying predictions on the EEG time series, and (4) saving high-quality images of the results. EPViz includes basic preprocessing operations, spectral feature extraction, and annotation editing. EPViz’s built-in anonymizer can facilitate the sharing of scalp EEG data between clinicians and engineers. EPViz fills a much needed gap in EEG visualization and may foster new research directions.

Results: EPViz can be used in a variety of clinical and research applications where the goal is to detect an event from EEG data. One natural domain is epileptic seizure detection, where EPViz can be used to view the seizure onset and offset predictions made by ML algorithms. Other potential applications are aura and non-epileptic event detection, both of which apply similar training and evaluation strategies. EPViz supports channel-wise predictions, which makes it a natural tool for seizure localization studies, where the goal is to identify a specific area of onset and track the seizure activity as it propagates. Finally, EPViz can overlay “predictions” contained in an auxiliary file, which the user can create manually based on clinical annotations or experimental conditions. Thus, it can be used even outside of the predictive analytics domain.

Conclusions: We have introduced EPViz, a user-friendly visualizer for EEG data. EPViz is a valuable tool for epilepsy researchers, allowing the results of predictive models to be overlaid on the EEG signal and on a topological head plot. Such feedback provides a mechanism to interpret “black-box” ML methods with respect to the data. EPViz can generate high-quality images, apply standard filtering, organize EEG signals, annotate EDF files, and even anonymize clinical data. EPViz is completely open-source and uses Python, the fastest-growing programming language for ML. EPViz is freely available for download at https://engineering.jhu.edu/nsa/links/.

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.