Multi-dimensional feature extraction of EEG signal and its application in stroke classification

Feature extraction based on EEG signals and construction of classification models using machine learning methods are key to intelligent assisted diagnosis of brain diseases (such as stroke classification). However, the quality of the extracted features directly affects the classification performance. This study proposes a multi-dimensional feature extraction method based on autocorrelation and complexity theory. It introduces an improved multifractal detrended fluctuation analysis (MFDFA) algorithm based on optimized empirical mode decomposition to extract high-quality autocorrelation features. In addition, we find that the ratio of fuzzy entropy between high-frequency band and low-frequency band of cerebral infarction signals is significantly lower than that of cerebral hemorrhage signals. On this basis, we propose a new complexity feature - fuzzy asymmetric index (FAI) based on constant Gaussian membership function. The study then integrates hierarchical fuzzy entropy, asymmetric entropy, and FAI to obtain complex fusion features. These extracted and fused features demonstrate excellent classification performance for differentiating cerebral hemorrhage and cerebral infarction. Using the random forest algorithm with a constant Gaussian membership function, the classification achieves an accuracy of 99.33%, precision of 100%, sensitivity of 98.57%, specificity of 100%, F1-score of 99.23%, and MCC of 98.73%. The proposed multi-dimensional features, combining autocorrelation and complexity characteristics, perform remarkably well in the classification of stroke EEG signals.