A Novel Feature Analysis Method for EEG Signal Classification of Attended and Unattended Audiovisual Integration

Our brains selectively attend to current task-relevant stimuli while ignore those sensory inputs irrelevant in plenty of information from different senses every day. The attention mechanism can help the brain select and filter effective information. Attention can affect single sensory information, as well as multiple sensory systems, in a top-down manner based on the goals and intentions of observers. Attention is an important factor affecting cognition, and can modulate various stages of the cognitive process. The comparison of attended versus unattended audiovisual multisensory EEGs showed some significant differences, providing the possibility for us to classify the cognitive process with or without attention involvement. The classification and recognition of cognitive processes with or without attention involvement is helpful for the diagnosis and recognition of mental disorders related to attention deficits. Here, we designed an audiovisual integration experiment using semantically congruent audiovisual stimuli to record the EEG signals associated with attended and unattended audiovisual stimuli. Then, statistical tests were used to detect the time and corresponding electrodes with significant differences between these two types EEGs of audiovisual integration. The peak amplitude and latency of these electrodes with significant differences were entered into SVM classifier as feature vector sets for training and testing. We further used the sequential forward searching (SFS) method to minimize the eigenvectors set. The result showed that the peak amplitude and latency of channel FP2, F1, FZ, F8, FC1, C5, AF3, F2, F4, FC5, FC3, FCZ, T7, F7, F3, F5, FC2, FC6, C1, AF4, F6, FT7, AF4, F6, FT7, FC4, C3, C6, T8, FT8 in time period of 260-300 milliseconds as the eigenvector reached the highest classification accuracy of 71.43%. This indicated that the difference between attended and unattended audiovisual integration in fronto-central area are the most significant, further indicating that the attention in the frontal lobe has the strongest modulation effect on audiovisual integration in this time range. In addition, a SVM classification model was obtained through a novel feature analysis method, which increased the classification accuracy to more than 70%, indicating that the evaluation model based on this method was effective. We hope that our work will provide effective help for objective diagnosis of attention-related cognitive impairment.