Multi-ganglion ANN based feature learning with application to P300-BCI signal classification

• Published in: Biomedical signal processing and control Vol. 18; pp. 127 - 137
• Main Authors: Gao, Wei, Guan, Jin-an, Gao, Junfeng, Zhou, Dao
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.04.2015
• Subjects: Imitating-Reading BCI; Multi-channel signal feature extraction; Multi-ganglion artificial neural network; Imitating-Reading BCI; Multi-ganglion artificial neural network; Multi-channel signal feature extraction
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2014.12.007

•The multi-ganglion ANN based feature learning (ANNFL) method is an unsupervised feature extraction method.•This method can find an effective feature representation automatically for single-trial P300 signal.•The ANNFL method reduces the training time of conventional three-layer auto-encoder and leads to better classification results in the P300-BCI paradigm of our study. The feature extraction of event-related potentials (ERPs) is a significant prerequisite for many types of P300-BCIs. In this paper, we proposed a multi-ganglion artificial neural network based feature learning (ANNFL) method to extract a deep feature structure of single-trial multi-channel ERP signals and improve classification accuracy. Five subjects took part in the Imitating-Reading ERP experiments. We recorded the target electroencephalography (EEG) samples (elicited by target stimuli) and non-target samples (elicited by non-target stimuli) for each subjects. Then we applied ANNFL method to extract the feature vectors and classified them by using support vector machine (SVM). The ANNFL method outperforms the principal component analysis (PCA) method and conventional three-layer auto-encoder, and then leads to higher classification accuracies of five subjects’ BCI signals than using the single-channel temporal features. ANNFL is an unsupervised feature learning method, which can automatically learn feature vector from EEG data and provide more effective feature representation than PCA method and single-channel temporal feature extraction method.