Classification of normal and epileptic seizures from EEG signals using Explainable Artificial Intelligence
machine learning, explainable AI, electroencephalography, epilepsy.
Epilepsy is a disease that affects 50 million people worldwide, one of its characteristics is the predisposition to sudden seizures. This fact negatively impacts the quality of life of those affected by the pathology. Therefore, many efforts have been employed in an attempt to predict epileptic seizures through the analysis of electroencephalogram (EEG) signals, using machine learning and deep learning techniques. In this work, a methodology is presented that analyzes the EEG signal in the time domain, in which features are calculated in time windows of 1s for later reduction of the number of features that are used in the training of a supervised classifier. The relevance of this method consists in reducing the number of EEG channels and the number of input vectors of the classifier, enabling the creation of a more concise model capable of achieving accuracy similar to the most robust models in the literature. For this, the SHapley Additive exPlanations method (SHAP Values) was used, which makes use of Explainable Artificial Intelligence, a technique that allows the interpretation of the characteristics that most contribute to the output of the model. With this method, an accuracy greater than 95% was obtained in the prediction of binary data (normal and seizure) from the University of Beirut Medical Center database, using only 5 channels and 6 input vectors. It is expected, therefore, that in future works it will be possible to further reduce the number of attributes using specific training models for each subject.