A Feature Fusion Network with Spatial-Temporal-Enhanced Strategy for the Motor Imagery of Force Intensity Variation

Ankai Ying^1,2,3Jinwang Lv^1,2,3Junchen Huang^1,2,3Tian Wang⁴Peixin Si^2,3Yi Zhang^2,3Jiyu Zhang^4*Guokun Zuo^1,2,3*Jialin Xu^1,2,3,5*

• ¹Cixi Institute of Biomedicine, Wenzhou Medical University, Cixi, China
• ²Ningbo Institute of Materials Technology and Engineering, Chinese Academy of Sciences (CAS), Ningbo, Zhejiang Province, China
• ³Ningbo Cixi Institute of Biomedical Engineering, Ningbo, Zhejiang Province, China
• ⁴Hangzhou RoboCT Technology Development Co., Ltd., Hangzhou, China, Hangzhou, China
• ⁵University of Chinese Academy of Sciences, Beijing, Beijing, China

Motor imagery (MI)-based brain-computer interfaces (BCI) offers promising applications in rehabilitation. Traditional force-based MI-BCI paradigms generally require subjects to imagine constant force during static or dynamic state. It is challenging to meet the demands of dynamic interaction with force intensity variation in MI-BCI systems. To address this gap, we designed a novel MI paradigm inspired by daily life, where subjects imagined variations in force intensity during dynamic unilateral upper-limb movements. In a single trial, the subjects were required to complete one of three combinations of force intensity variations: large-to-small, large-to-medium, or medium-to-small. During the execution of this paradigm, electroencephalography (EEG) features exhibit dynamic coupling, with subtle variations in intensity, timing, frequency coverage, and spatial distribution, as the force intensity imagined by the subjects changed. To recognize these fine-grained features, we propose a feature fusion network with a spatial-temporal-enhanced strategy and an information reconstruction (FN-SSIR) algorithm. This model combines a multi-scale spatial-temporal convolution module with a spatial-temporal-enhanced strategy, a convolutional auto-encoder for information reconstruction, and a long short-term memory with self-attention, enabling the comprehensive extraction and fusion of EEG features across fine-grained time-frequency variations and dynamic spatial-temporal patterns. The accuracy rate of the FN-SSIR was 86.7±6.6%, which is 8.5% higher than that of the baseline algorithms. These findings highlight the potential of this paradigm and algorithm for advancing MI-BCI systems in rehabilitation training based on dynamic force interactions.