Abstracts

Rationale: Long-term video EEG monitoring (vEEG) is the gold standard for determining if an individual is having seizures. However, it is exceedingly time consuming to examine up to 24 hours of vEEG every day for each patient. Further, seizure identification is an intrinsically complex task and research shows there is considerable variability in human performance on this critical task.Recent advances in machine learning (ML) methods are making it possible to use computers to classify complex data sets, which may make it feasible to identify seizures with minimal human interaction with the EEG recording. We present a case study for detecting seizures that makes use of ML algorithms to speed up the EEG reading process. The goals were lightweight data pipelining to ensure model speed and classification model accuracy in the low-to-mid 80s range. Methods: The patient was an eight-year-old female with epilepsy since age four. She presented with frequent absence seizures lasting from 5 – 10 seconds each and generalized tonic clonic seizures. The patient was monitored for 6 days in the EMU at Dell Children’s Medical Center. vEEG was recorded continuously with a 19-channel montage using the International 10-20 system and seizures were identified by an experienced epileptologist.There were 16 seizures used in this study, totaling 343 seconds. The identified seizures were made into 4 second clips. In addition, an equal number of 4 second clips were made from the patient’s interictal EEG data using sleep and wake periods. The spectral power of each channel of each clip was computed using the SciPy toolbox. The spectral power of the EEG from .5 – 40 Hz. was used in the classifier.The model was built in Python 3.6 using the Keras API. The platform used was Tensorflow-GPU. NVidia's cuDNN libraries were utilized to accelerate training. The ML architecture used was a simple recurrent neural network (RNN) with a SimpleRNN layer, followed by densely connected layers into a sigmoid output. To improve accuracy, gated recurrent units (GRU) were used in the RNN.The classification model was built using supervised learning of randomly selected clips from seizure and interictal recordings. The test data for the model was selected from different days than the training data. Results: The self-reported classification accuracy of the best model was 81.2%. The calculated sensitivity of the model to detect seizures was 78.6% and the specificity was 90.4%. The total accuracy of clips correctly classified was 83.9%. Conclusions: High model accuracy was attained with very good sensitivity and specificity. The total accuracy of clips correctly classified met our pre-established seizure detection goal. The minimal signal processing used in this study demonstrates the feasibility of examining long-term EEG recordings quickly to mark probable seizures for review by physicians.The major limitation of this study is that it involved only a single subject, yet it shows the potential of ML in the detection of seizures in recorded EEG as a tool to improve patient care. Additional discussion will show how this method can be generalized and applied in a clinical EMU setting. Funding: No funding was received.