EEG signal biometric extraction to predict stroke severity

Quantitative measures of specific electroencephalogram features are valuable for clinical applications to evaluate stroke characteristics. The EEG signals of 48 subjects were collected under the FIST task to (i) identify the most accurate EEG index that distinguishes different stroke severity levels and assess them using machine learning classifiers; (ii) develop a model to rapidly predict stroke severity before admission. We calculated the relative power and pairwise-derived brain symmetry index over different frequency bands (1–3 Hz, 4–7 Hz, 8–13 Hz, 14–30 Hz), delta/alpha ratio (DAR), (delta + theta)/ (alpha + beta) ratio (DTABR), and Shannon entropy. An analysis of variance was performed to screen the features. The filtered features were input into different classifiers to verify the most accurate EEG index and classifier performance metrics. The relative power of the theta band, DAR, DTABR, and Shannon entropy of the delta band were identified as the four factors with the most significant differences. The classification accuracies of Support Vector Machine, K-Nearest Neighbors (KNN), Decision Tree, and Naive Bayes were 82.57%, 89.82%, 80.35%, and 79.32%, respectively. The results indicate that the relative alpha power, DAR, DTABR, and Shannon entropy of the delta band contain unique information indicative of stroke severity. The KNN classifier demonstrates the highest overall classification accuracy among the four classifiers.