StrokeSight: A Novel EEG-Based Diagnostic System for Strokes Using Spectral Analysis and Deep Learning

Abstract

A stroke is defined as a neurologic deficit arising from an interruption in blood supply to the brain. According to the World Health Organization, over 15 million people suffer from strokes annually, of which almost 70% die or are permanently disabled. Effective treatment must be administered within one hour to prevent irreversible brain damage. Unfortunately, the current gold standards for diagnosis, CT and MRI, are time-consuming, expensive, and immobile. Electroencephalograms reveal biomarkers of strokes while being inexpensive and available for remote use, but no system exists that utilizes them for this purpose. To address this issue, we created StrokeSight, a novel, open-source web application that automatically provides a full diagnosis and visualization of ischemic and hemorrhagic strokes in under 50 seconds using 60-second electroencephalograms. We first calculated the averaged power spectral densities for 132, 60-second electroencephalogram readings, which we then used to train three deep neural networks that respectively predict a stroke type (control/ischemic/hemorrhagic), location (left/right hemisphere), and severity (small/large) with accuracies of 97.5%, 94.4%, and 100%. StrokeSight also implements a novel process to visualize spectral abnormalities caused by strokes. Azimuthal equidistant projection and multivariate spline interpolation are used to reshape 3D electrodes onto a head-shaped 2D plane and then a contour map of each frequency band power is created, allowing neurologists to quickly and accurately interpret electroencephalogram data. StrokeSight could act as a revolutionary solution for stroke care that drastically improves the speed, cost efficiency, and accessibility of stroke diagnosis while allowing for personalized treatment and interpretation.