An EEG-Based Diagnostic Framework for Strokes Using Spectral Analysis and Deep Learning

Strokes affect over 15 million people annually, and treatment must be given within one hour to prevent brain damage. CT and MRI scans are used for stroke diagnosis but are time-consuming and expensive. Electroencephalograms (EEG) solve both problems but are not used due to their complexity. This paper presents a novel application for the automated detection of ischemic and hemorrhagic strokes using EEGs. We propose using the averaged power spectral density to extract important features. We use deep neural networks to respectively detect and classify the stroke type, location, and severity with accuracies of 97.5%, 94.4%, and 99%. Additionally, to allow for easier EEG interpretation and detection of abnormalities, we use azimuthal projection and spline interpolation to reshape 3D electrodes onto 2D contour maps showing the power of frequency bands around the brain. This research could represent a step towards increasing the speed and decreasing the cost of comprehensive stroke diagnosis.