فصلنامه مهندسی پزشکی زیستی
سال شانزدهم شماره 1 (بهار 1401)

Visual attention as a cognitive factor plays a significant role in the processing of higher-order mental information that happens in the brain and affects brain activity in various areas of the visual cortex. Among the various recording systems, local field potentials, due to their stability, robustness, and frequency content have received interest in brain structure and cognitive processing research, as well as brain-computer interface (BCI) systems. Hence, the extraction and interpretation of information from local field potential (LFP) signals during visual attention has been considered to control cognitive systems. Cross-frequency coupling (CFC) as one of the information encoding strategies in the brain plays a functional role in perception, working memory, and visual attention tasks. However, the role of CFC as informative features for spatial attention decoding has not been adequately investigated. This paper aims to examine spatial attention decoding using LFP signals recorded from the monkey middle temporal area (MT). For this purpose, phase-phase and phase-amplitude coupling features and machine learning algorithms have been employed. The results show that the highest decoding performance was achieved by applying selected optimal features and the support vector machine classifier (90.36%). Moreover, among the selected features, gamma-delta, gamma-alpha, and beta-delta coupling contain the most cognitive information and the most effective features to improve the decoding performance of spatial attention in the visual system. Generally, the results suggest that cross-frequency coupling of LFP signals contains significant information in spatial attention tasks, and can be used as a suitable alternative to the time-frequency features of brain signals in cognitive BCI systems.

The human brain is one of the most complex and heterogeneous networks, and brain signals contain a lot of information, so researchers in this field are always looking for proper solutions to select meaningful features and reduce the dimension of this information appropriately to lead to better classification. Two of the new tools for brain signal processing are Graph Signal Processing (GSP) and Meta-heuristic and Evolutionary methods. In this paper, a geometric structure and a mixed structure are considered for the brain graph and the weights of the edges in the mixed structure are calculated by a combination of two measures: geometric distance and correlation. To reduce the graph dimension, the weighted degree metric and a combination of the Kron reduction method and Graph Fourier Transform (KG) are used to properly preserve the information of all vertices of the graph into the selected vertices. Feature extraction is performed by Ledoit-Wolf shrinkage estimation and Tangent Space Mapping (TSM) method. For dimension reduction of extracted features, Principal Component Analysis (PCA) method and feature selection based on Differential Evolution (DE) are used. The selected features are given to several well-known machine learning classifiers. To evaluate the performance of the proposed method, dataset IVa from BCI Competition III has been used. The results show that the average classification accuracy of the proposed KG-PCA method with SVM-RBF and DT classifiers, in the structural graph and the functional-structural graph, is higher than the G-PCA method expressed in previous studies, and the DT classifier has achieved an average accuracy of 91.15 ± 1.17. Also, according to the obtained results, the performance of the proposed KG-DE method has been better compared to KG-PCA and in the best case, the average accuracy of the SVM-RBF classifier is equal to 95.50 ± 1.27.