An Adaptive-Load Classifier based on Cognitive Load Similarity

• Published in: IEEE sensors journal Vol. 23; no. 18; p. 1
• Main Authors: Song, Zhiyong, Li, Mengfan, Wu, Lingyu, Wu, Yuwei, Xu, Guizhi, Lin, Fang, Liao, Wenzhe
• Format: Journal Article
• Language: English
• Published: New York IEEE 15.09.2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: adaptive-load classifier; Background noise; brain load similarity; Brain research; Brain-computer interface; Channels; Classifiers; Cognitive load; Correlation coefficients; Electroencephalography; Entropy; Human-computer interface; Load matching; Sensors; Similarity; Task analysis; Training; Visualization
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2023.3299086

Environmental noise leads to changes of the people's brain load, making it necessary for the brain-computer interface (BCI) to adapt to the load changes. Our previous research found BCI performance is influenced by the similarity between the current test load and the past train load. Based on the similarity of the current load to three past loads, this research proposes an adaptive-load classifier to dynamically adapts to the current brain load by adjusting the weights of three load specific classifiers by brain load similarity. Pearson correlation coefficient (PCC) and sample entropy (SE) are separately applied to match the similarity of the brain loads. Then the adaptive-load classifier transforms the similarities to the weights to fuse the load specific classifiers. Ten subjects participated in this study, using 28-channel and 3-channel EEG signals to test the classifier. The PCC and SE reach higher accuracies and ITRs than the load specific classifier with an improvement of 31.07% ( P =8.63×10-5) and 7.34% ( P =0.05), 32.15 bits/min ( P =0.02) and 7.08 bits/min ( P =0.24) in 28 channels, and reach the accuracies of 99.17% and 91.67%, and the ITR of 38.05 bits/min and 26.14 bits/min in 3 channels. The adaptive-load classifier could adapt to different loads by matching the similarity of the current data and past loads with a large amount or a small number of channels, which is practical for the bCi application in the real environment.