The brain-computer interface (BCI) holds immense potential to enhance independence for individuals with motor impairments in daily life. However, decoding neural signals related to movement parameters remains a significant barrier to seamless real-world control. In this paper, we propose DeCLFPNet, a parameter-efficient deep neural network designed for high-accuracy decoding of continuous movement and applied force from local field potential (LFP) signals. DeCLFPNet Optimizes the integration of a convolutional block to extract optimal features and a recurrent block to estimate movement parameters, utilizing multiple techniques to boost decoding performance while maintaining a relatively small number of parameters. We evaluated DeCLFPNet on two datasets: the Rats’ LFP dataset for applied force decoding and a publicly available Monkey's LFP dataset for 2D movement decoding. DeCLFPNet outperformed conventional approaches, including linear models and other deep networks, on both datasets, achieving average correlation coefficients (r) of 0.88 for position and 0.89 for force, with coefficients of determination (R 2 ) of 0.77 and 0.76, respectively. Notably, DeCLFPNet demonstrates superior performance compared to leading methods on the Rats’ LFP dataset, with R 2 values of 0.76 and 0.49 for within-subject and cross-subject decoding scenarios, respectively. Our approach enhances feature extraction from LFP signals, improving generalization across subjects and datasets, and provides insights into the physiological relationships between LFP signals and movement parameters, illuminating neural dynamics of motor control.
Bina et al. (Wed,) studied this question.