Hybrid Stacking Model for Automatic Epileptic Seizure Detection Using Electroencephalogram Signals

This paper proposes a hybrid machine learning model based on stacking technique. The proposed model is intended for automatic detection of pathologies associated with epilepsy using an electroencephalogram signal (EEG). This model has a two-level architecture. The first level of the model is a set of independent base classifiers. The second level is the final classifier (meta-model). The set of basic classifiers includes Decision Tree (DT), Random Forest (RF) and Support Vector Classification with linear kernel (SVC linear). A Support Vector Classification with polynomial kernel (SVC polynomial) is used as a meta-model. This approach, which combines various classifiers, makes it possible to achieve improved performance metrics of the model compared to the basic models included in its composition. In addition to the frequency and statistical characteristics of the signal, methods based on the concept of fractals and chaos from the field of nonlinear dynamics were used to extract informative features. Computational experiments have shown an increase in the Accuracy and Specificity of the proposed model compared to the basic classifiers included in its composition, as well as a better Sensitivity value compared to models of other authors on the open dataset “TUH EEG (TUSZ)”.