Comparison of Machine Learning Algorithms for Seizure Detection on EEG

This project is sought to compare the performance of different supervised machine-learning algorithms in the detection of seizures in electroencephalograms (EEG). To contrast the results of the algorithms metrics of performance are used, this allows an adequate comparison to be made. The EEG signals analyzed belong to the public database: “CHB-MIT Scalp EEG Database” which contains EEG records of 24 pediatric patients presenting with seizures. For its analysis, it is decided to detect seizures in periods of 5 min. To have descriptive data from this period, it is necessary to analyze the signals. They are processed using the fast Fourier transform to obtain the frequencies that put together brain activity, they are classified according to the frequency of the different brain waves, and their energy levels are measured. Consecutively, the models obtained are evaluated by adjusting the different hyperparameters of the selected algorithms in order to achieve the best performances in each one. Finally, it is concluded that the machine-learning algorithms that most accurately predict seizures are GBM, kNN, and neural networks.