BiCurNet: Premovement EEG-Based Neural Decoder for Biceps Curl Trajectory Estimation

Abstract

Kinematic parameter (KP) estimation from early electroencephalogram (EEG) signals is essential for positive augmentation using wearable robots. However, surface EEG-based early KP estimation studies are sparse in the literature. In this study, simultaneous surface EEG and kinematics data of five participants is collected during the biceps-curl motor task. The feasibility of early estimation of KPs is demonstrated using brain source imaging (BSI). Discrete wavelet transform (DWT) is utilized for subband extraction from preprocessed EEG signals. Further, spherical and head harmonics domain features are extracted from subbands of the EEG signals. A deep-learning-based decoding model, BiCurNet, is proposed for early KP estimation using spatial and harmonics domain EEG features during the biceps-curl task. The proposed model utilizes lightweight architecture with depthwise separable convolution layers and a customized attention module (CAM). The best Pearson correlation coefficient (PCC) between the estimated and actual trajectory of 0.7 is achieved when combined EEG features (spatial and harmonics domain) in the delta band are utilized. Intra- and intersubject performance analyses are performed to evaluate the subject-adaptability of the proposed decoding model. The performance of the proposed BiCurNet is compared with the existing multilinear regression (mLR) counterpart. The robustness of the proposed model is additionally illustrated using an ablation study. The robust performance and lightweight architecture will facilitate the real-time implementation of the model with deployment on a microcontroller to control BCI-based wearable robots.