Salient feature extraction for EEG pathology detection

Inhalt:
Electroencephalogram (EEG) is an economical and convenient auxiliary test to aid in diagnosis and analysis of neurological diseases. Nevertheless, due to the increasing morbidity of brain disorders and the limitations of direct visual inspection by trained neurophysiologists, i.e., being arbitrary and time-consuming, developing a computer-aided diagnosis system that could provide pre-diagnostic screening would be of great significance. This paper proposes an efficient method based on aggregation technique for EEG pathology detection. First, we adopt wavelet packet decomposition to break down the EEG signals into various frequency components, and some statistical features are calculated from each selected component. Subsequently, a novel and efficient technique is presented to decrease the dimensionality of the constructed feature space without degrading the feature quality. Finally, two different ensemble learning models are used to classify the extracted salient features into normal or abnormal classes. In a widely accepted EEG benchmark dataset, the categorical boosting-based classification method achieved accuracy and G-mean of 88.76% and 88.18%, respectively, comparing favorably to other recently published methods in performance. The results suggest that the highly compact and discriminative features could be effectively taken from raw EEG signals using the salient feature extraction technique and may pave the way for the development of a reliable and rapid abnormal EEG detection method.