Abstracts

Rationale: The analysis of Video-EEG data is a necessary step in rodent studies evaluating anti-epileptic therapies or epileptogenesis mechanisms. Manual seizure review and annotation of this data is time consuming for researchers. Given the plethora of differing setups for gathering data and the rapid growth of machine learning and deep learning modules for seizure detection that exist, there is a need for a system that allows for easy research integration with minimal barriers to adoption for researchers with limited coding-expertise. Our goal, through EEnGine (an EEG analysis “engine”), is to enable an intuitive graphical-user interface (GUI) platform for seizure detection and prediction.


Methods: EEnGine’s modular design allows researchers to utilize the platform regardless of coding level. The modular architecture consists of 1) a compatibility layer to translate file formats from various EEG acquisition systems into standard EEG objects for processing, 2) an optional feature extraction layer to extract machine-learning inputs, 3) user-configured training modules that allow researchers to define their own machine learning or deep learning modules, and 4) a classification system that utilizes trained models to classify previously unseen data. The GUI joins this coding architecture together such that researchers can simply select between their preferred modules to either train on, or classify, their data without having to directly interact with code. The GUI is as simple to install and use as any other app, providing EEG trace visualization of multiple channels. For our purposes, we created novel seizure detection machine learning models in the context of chronic, invasive EEG recording in an intra-amygdala kainic acid mouse model, representative of temporal lobe epilepsy. The machine learning models were evaluated on EEG recordings including 90 days of 24/7 monitoring on this mouse model (West et al., Exp Neurol, 2022; 349:113954).

Results: Modules were trained in EEnGine for specific use cases such as online and post-hoc analysis. Post-hoc analysis emphasizes accuracy, so a more complex module was developed with accuracies in the range of human annotators. Online analysis requires low latency and operational robustness, so less computationally expensive features were used in a simpler module emphasizing seizure onset detection. EEnGine was found to meet usability requirements through evaluation with set tasks and a quantized scoring system.

Conclusions: EEnGine and the novel trained machine learning models presented will further enable seizure detection and prediction methods. These methods will streamline Video-EEG rodent epilepsy studies that are important for drug and therapy development. By using EEnGine, researchers will be able to reduce time spent on developing in-house solutions. A future direction will be to enable collaboration on clinical data, allowing researchers to collaborate between rodent and clinical studies.

Funding: Australian Government Research Training Program Scholarship

Dr. Anderson's start-up package (School of Biomedical Engineering, The University of Sydney)