A Pilot Study On Detecting Cognitive Load Using Kolmogorov-Arnold Networks With EEG Signals

Cognitive load detection is a critical part of cognitive state research, aiding in the understanding of the brain's response to stimuli in the cognitive process. Electroencephalogram (EEG) is a valuable tool for viewing brain states under varying cognitive loads. This study proposes a cognitive load detection approach based on Kolmogorov-Arnold Networks (KANs), a novel neural network architecture, utilizing multi-channel EEG signals. The five different types of features (raw EEG signals, temporal statistical features, time-frequency features, entropy features, and deep features) are used as inputs to explore the efficiency of three KAN-based detection models (KAN, fast-KAN, & Cheby-KAN). Furthermore, this study investigates the effect of different channels on the detection performance using different feature inputs. The experiments are conducted using a 19-channel EEG dataset from a mental arithmetic task. The results demonstrate the proposed KAN-based models outperform existing state-of-the-art deep learning-based methods (e.g., EEGNet & CNN-BiLSTM), achieving an accuracy of 98.26% and an F1-score of 98.30%. KANs obtain better performance than MLPs as classifiers at the decision-making layer in this EEG classification task. Moreover, the later channel selection experiments reveal that the selected subsets of channels hold the potential to achieve similar performance to the full set of channels.