Epileptic Seizure Detection via EEG using Tree-based Pipeline Optimization Tool

Abstract

The electroencephalogram (EEG) signals being a recording of the electrical activity of the brain provides valuable information in the analysis of its function and disorder. Epilepsy is a brain disorder characterized by uncontrolled excessive activity. Repeated abnormal disturbance or seizure causes epilepsy. Hence, EEG signals act as a diagnostic tool for epilepsy. An approach based on tree-based pipeline optimization tool (TPOT) is presented for classification of EEG signals as either seizure or normal activity in the brain. A binarized dataset with sampled signal levels and the corresponding class is subjected to a genetic approach for acquiring an optimized predictive model. In TPOT, the tedious process involved in machine learning being repeatedly performed until arriving at the best solution is automated using genetic algorithm, i.e., evaluate-select-crossover-mutate is repeated to tune the pipeline. In the settings used in this paper, there are 90 pipeline configurations for evaluation for which around 450 models are fitted and evaluated against the training data in one grid search. The best pipeline is the one with the highest cross-validation score in the run at 95.94%. The test accuracy is at 95.27% which is just a little lower than the cross-validation score. The predictive model consists of pre-processing steps Maximum Absolute Scaler and Function Transformer which is utilized by a Gaussian Naive Bayes classifier. The system is trained and tested for epileptic seizure detection using raw EEG signals. The optimized features and predictor obtained via TPOT resulted to a high-performance accuracy for epileptic seizure detection.