SCVCNet: Sliding Cross-Vector Convolution Network for Task-Independent Cognitive Workload Recognition With EEG Measurement

This article presents a generic approach for measuring cognitive workload by exploiting common electroencephalogram (EEG) patterns. We propose a neural network called SCVCNet, which eliminates task- and individual-set-related interferences in the EEG by analyzing its detailed spatial-frequency structures. The SCVCNet utilizes a sliding cross-vector convolution (SCVC). It adopts two input layers representing theta and alpha power at fine-grained frequency points, with each node corresponding to an EEG electrode. For each frequency point, the inputs from the two bands are arranged in a cross shape. A kernel matrix is then applied to the power values at the intersection of the electrode locations, focusing on both the central and neighboring electrodes. Next, we introduce an interfrequency-point feature integration (IFPFI) module to fuse the SCVC feature maps. We validate its performance using three databases, each consisting of distinct tasks performed by independent participant groups. The average accuracy (ACC) (68.1% and 62.3%) and F1 score (67.4% and 60.8%) achieved in two different validation paradigms show partially higher performance than the previous works. All features and proposed algorithms are publicly available.