A Fusion Feature for Enhancing the Performance of Classification in Working Memory Load with Single-Trial Detection.

In traditional brain-computer interfaces (BCIs), using only a certain type of feature or a simple mixture of different features cannot meet the requirements for high performance in classification. Moreover, a simple mixture of various features might lead to information redundancies and thus increase the computational complexity. In this paper, we studied the feasibility of integrating two kinds of features, which showed opposite variation trends as the memory load levels increase, into a single fusion feature. We also proposed a feature fusion framework based on non-invasive electroencephalography to classify the memory load levels and estimate the workload for a series of challenging working memory (WM) tasks (involving delayed match-to-sample tasks) on a single-trial basis. A novel fusion feature called spectral entropy/Lempel-Ziv complexity (SEn/LZC) was proposed to classify three memory load levels. The results showed that the generalization of the support vector machine (SVM) with SEn/LZC was significantly higher than the generalization of an SVM with four other types of feature, namely SEn, LZC, SEn&LZC and LZC/SEn. The findings suggested that the proposed fusion feature could act as a biomarker to successfully distinguish different load levels and that the constructed framework could achieve consistency between optimal cognitive performance and fusion features. In addition, the proposed fusion framework could provide a new method of successfully promoting the classification generalization of BCI and implicitly evaluating the mental workload.