ECoG Dual Context Network for Brain Machine Interfaces

• Published in: IEEE access Vol. 13; pp. 99252 - 99265
• Main Authors: Nagashima, Shunya, Kaneda, Kanta, Wada, Yuiga, Iida, Tsumugi, Taguchi, Misa, Hirata, Masayuki, Sugiura, Komei
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Amyotrophic lateral sclerosis; BMI; Brain modeling; Classification; Communication; Conformer; Context; Convolution; ECoG; Electroencephalography; Feature extraction; Grasping; Hands; Keyboards; Man-machine interfaces; Motors; Movement; Muscular dystrophy; retention; Speech; Support systems; Time series analysis; Transformers
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2025.3577407

Speech impairments in conditions such as amyotrophic lateral sclerosis (ALS) and muscular dystrophy significantly limit communication with patients, profoundly impacting their daily life and social participation. Communication support through decoding technology based on electrocorticography (ECoG) is thus critically important to assist the day-to-day communication of patients with speech impairments. This study proposes an electrocorticography (ECoG)-based communication support system to enhance communication for individuals with speech impairments, such as those caused by amyotrophic lateral sclerosis (ALS) or muscular dystrophy. In our system, we classify ECoG signals recorded during movement attempt tasks involving the upper limbs, such as hand grasping. This classification enables the control of a virtual keyboard, allowing for text-based communication. Specifically, predicted movement attempts are associated with keyboard functions, such as moving the cursor or pressing the select button. In this study, we propose the ECoG Dual Context Network (ECoG DCNet), designed to efficiently embed information between electrodes while preserving the length of temporal sequences to capture long-range dependencies. To validate our approach, we used a clinical dataset from a single paralyzed subject with ALS, focusing on the ECoG recorded during movement attempt tasks. The proposed method outperformed baseline methods in terms of classification accuracy. Notably, our method achieved an accuracy of 75.3%, outperforming the widely-used EEGNet by 30%. Moreover, Gradient-weighted Class Activation Mapping visualizations suggested that the proposed method focuses on features that could be neurophysiologically relevant.