Background: Epilepsy is a neurologic disease characterized by seizures which occur due to sudden and synchronized bursts of excessive electrical energy in the brain. An electroencephalogram, or EEG, can detect seizures in real time but requires trained medical expertise for extended periods of time. The main objective of this research was to devise a more efficient method (SeizureSeeker) for analyzing EEG data using machine learning algorithms that allows for complex data processing and can automatically distinguish between normal EEG signal and epileptic seizures.
Methods and Study Design: An open access EEG dataset, containing pre-identified records of 500 patients, was used. Seizure activity was designated as a simple binary 1 or 0, where 1 indicated a seizure and 0 indicated no seizure. The database was then partitioned into two randomly assigned groups, a training set of 80% of the data and a testing set containing the remaining 20%. The study compared 3 different classification algorithms: Logistic Regression, Support Vector Machines (SVM), and Long Short-Term Memory (LSTM). All models were fitted using existing software from Python libraries and the Orange data mining application.
Results: Logistic Regression had poor accuracy, but SVM achieved impressive results with an overall accuracy of 94%. LSTM is a more complex algorithm based on recurrent neural networks and generated near perfect classification results with an accuracy of 99%.
Conclusions: The memory property of the LSTM model makes it an ideal choice for the time series EEG data. The LSTM results proved the efficacy of the machine learning model to automatically detect seizure activity in EEG data. Models such as SeizureSeeker can be developed to reach more timely diagnoses of seizures and can be used where access to specialized medical expertise is especially limited.

PDF (Download Full text)