Steady-State Visually Evoked Magnetic Signal Classification and BCI Evaluation Based on a Convolutional Neural Network

• Published in: IEEE access Vol. 13; pp. 68622 - 68631
• Main Authors: Wei, Yutong, Zhao, Fudan, Zhao, Fengwen, Zheng, Shiqiang, Ye, Chaofeng, Liu, Liangyu
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Artificial intelligence; Artificial neural networks; Bio-magnetic field measurement; Brain; brain-computer interference; convolutional neural network; Convolutional neural networks; Data augmentation; Games; Magnetic field measurement; Magnetic fields; Magnetic shielding; Magnetic signals; Neural networks; Optical imaging; Optical sensors; Optical variables measurement; optically pumped magnetometer; Sensitivity; Signal classification; Statistical analysis; Steady state
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2024.3524397

The steady-state visually evoked magnetic field (SSVEF) is a promising modality in brain-computer interference (BCI), which has the advantages of being non-invasive and non-contact. The combination of optically pumped magnetometers (OPM) and artificial intelligence technology makes SSVEF measurements more portable, accurate, and cost-effective. This paper examines the distribution of the human brain visually evoked magnetic field experimentally and then presents an SSVEF measurement system based on an OPM. A three-block temporal convolutional neural network (3B-TCN) is developed to classify brain magnetic signals. A data augmentation method based on statistical analysis of SSVEF signals is proposed, which generates 30,000 sets of data to train the 3B-TCN. The SSVEF signal classification accuracies of the 3B-TCN network are 96.61%, 92.36%, and 86.75% for 10 s, 5 s, and 2 s time length data, respectively. The impact of visually fatigued states on BCI is studied. The accuracy of controlling the character in the game is above 90% when the subjects are in a normal state, but it decreases considerably when the subjects are visually fatigued. The experimental results demonstrate the feasibility of realizing BCI using an OPM sensor and a convolutional neural network.