EEG-Based Multiclass Workload Identification Using Feature Fusion and Selection

The effectiveness of workload identification is one of the critical aspects in a monitoring instrument of mental state. In this field, the workload is usually recognized as binary classes. There are scarce studies toward multiclass workload identification because the challenge of the success of workload identification is much tough, even though one more workload class is added. Besides, most of the existing studies only utilized spectral power features from individual channels but ignoring abundant interchannel features that represent the interactions between brain regions. In this study, we utilized features representing intrachannel information and interchannel information to classify multiple classes of workload based on an electroencephalogram. We comprehensively compared each category of features contributing to workload identification and elucidated the roles of feature fusion and feature selection for the workload identification. The results demonstrated that feature combination (83.12% in terms of accuracy) enhanced the classification performance compared with individual feature categories (i.e., band power features, 75.90%, and connection features, 81.72%, in terms of accuracy). With the F-score feature selection, the classification accuracy was further increased to 83.47%. When the features of graph metric were fused, the accuracy was reached to 84.34%. Our study provided comprehensive performance comparisons between methods and feature categories for the multiclass workload identification and demonstrated that feature selection and fusion played an important role in the enhancement of workload identification. These results could facilitate further studies of multiclass workload identification and practical application of workload identification.

Automated summary

This study classified multiple classes of workload based on EEG features, finding that feature fusion and feature selection played important roles in enhancing workload identification accuracy. Feature combination improved classification performance, with the highest accuracy achieved when graph metric features were fused.