Abstracts

Rationale: Seizure identification is a complex labor- and time-intensive task. When reviewing long-term EEG for seizures, it takes skilled technicians to review and edit the EEG record for physician review. Recent advances in machine learning (ML) methods and computers are making it feasable to classify complex data sets, which may make it possible to identify seizures with minimal human interaction. However, there has been limited success in these efforts.That lack of success is partly due to the complexity of the problem and the difficulty in representing that complexity so the ML model can identify associations and discriminant functions for accurate classification across a broad range of patients and seizure types. The complexity can be addressed in different ways, but the methods can all be grouped into two classes: simplifying the input data or building more sophisticated models.The first method, Feature Engineering, involves extracting features from the EEG data and applying machine learning algorithms to classify EEG as seizure or non-seizure. Any number of features, such as frequency band spectral power, coherence, laterality difference, etc., can be computed from the EEG and used to classify EEG clips.The second method, Feature Detection, involves applying multiple layers of sophisticated filters, combined with non-linear forward- and backward-propagating recurrent neural networks to build a mathematical model that can classify EEG data into seizure or non-seizure categories.  Methods: EEG data was clipped from the long-term monitoring of 78 pediatric patients who were evaluated at the EMU at Dell Children's Medical Center in Austin, TX. Patients ranged from 2 to 21 years of age. There were 37 female patients. The EEG was collected using the International 10-20 system and a 19-channel montage was selected for this study. There were 505 seizures identified by experienced epileptologists and a roughly equal amount of time was clipped from non-seizure periods.For feature engineering, Matlab was used to calculate the spectral power density in different frequency bands, the coherence between channels, and a difference score between left and right hemispheres, among others. These measures were computed for each channel and hemisphere. The ML model was built in Python 3.6 using the Keras API and the Tensorflow-GPU platform. The ML architecture was a recurrent neural network followed by 2 densely connected layers and sigmoid output.For feature detection, the EEG was low-pass filtered to 100 Hz and resampled to 256 Hz. The ML architecture was a series of convolutional neural networks (CNN), long short-term memory, and densely connected layers and sigmoid output.  Results: Multiple models were computed for both the feature engineering and feature detection methods. The best sensitivity performance for seizure detection was found using a series of four CNN and 2 densely connected layers for Feature Detection. The best performance was 67% accuracy classifying seizures in a test data set (not used during model training) and the overall accuracy for classifying both seizure and non-seizure clips was 80%. Using the Feature Engineering method reached a peak sensitivity of 53% accuracy for seizure detection and an overall accuracy of 64%. Conclusions: While both methods perform above chance for seizure detection and overall accuracy, neither method achieved the necessary accuracy for clinical utility. This suggests that we are still inadequately addressing the complexity of the problem.The discussion will explore how different methods can be combined to increase ML model performance and improve the generalizability.  Funding: No funding