ST-Topo GAN: A Motor EEG-to-EMG Decoding Model Matched to Wrist Movement Complexity
Authors: Ye Sun, Mingxuan Qu, Jing Wang, Dezhong Yao, Gang Liu
Organizations: School of Electrical and Information Engineering, Zhengzhou University, Zhengzhou 450001, China. · HumVerse (Zhengzhou) Technologies Co., Ltd., Zhengzhou 450001, China. · Henan Provincial Key Laboratory of Brain Science and Brain-Computer Interface Technology, Henan, China. · School of Mechanical Engineering, Xi’an Jiaotong University, Xi’an 710049, China. · Clinical Hospital of Chengdu Brain Science Institute, MOE Key Laboratory for NeuroInformation, Brain-Apparatus Communication Institute, University of Electronic Science and Technology of China, Chengdu, China. · Research Unit of NeuroInformation, 2019RU035, Chinese Academy of Medical Sciences, Chengdu 611731, China.
The wrist plays a critical role in upper-limb function by enabling precise hand positioning, force regulation, and object manipulation. Continuous brain--muscle interfaces (BMIs) offer a promising approach for motor restoration by decoding neural activity into muscle activation signals. However, existing EEG-to-EMG models have mainly been developed for tasks with relatively stable muscle synergies and may be less effective for the heterogeneous and weakly coupled neuromuscular organisation involved in wrist movements. This paper proposes ST-Topo GAN, a Spatial--Temporal Topological Generative Adversarial Network for continuous EEG-to-EMG decoding of wrist movements. The framework integrates multi-band EEG representation, sensorimotor cortical topology modelling, and conditional adversarial learning to reconstruct multi-channel iEMG activation. The model was evaluated through cross-task comparison, wrist EEG-to-iEMG decoding, and ablation experiments. Compared with the WAY-EEG-GAL grasp-and-lift dataset, the wrist dataset exhibited lower inter-muscle activation similarity and greater decoding difficulty for conventional models. ST-Topo GAN achieved an average PCC of 0.4436 on the wrist dataset, outperforming all evaluated baselines, while the ablation study confirmed the contribution of its key components. These results support the effectiveness of ST-Topo GAN for continuous wrist EEG-to-iEMG decoding.
Figures & tables
Fig. 1: Overview of the proposed ST-Topo GAN framework for continuous wrist EEG-to-iEMG decoding. The proposed framework reconstructs continuous multi-channel iEMG signals through conditional adversarial learning, where generated and real iEMG patterns are compared to improve the physiological plausibility of decoded muscle activation.
Fig. 2: Experimental paradigm of the self-collected wrist movement dataset. Participants performed wrist flexion and extension movements following visual cues. Each trial consisted of a 5 s movement period and a 5 s rest period, with repeated trials conducted across three experimental blocks.
Fig. 3: Illustration of EMG preprocessing and windowed muscle activation extraction. Each EMG channel was segmented into 500-ms windows, full-wave rectified, and summed to obtain the continuous iEMG activation target.
Fig. 4: Architecture of the proposed wrist-oriented ST-Topo GAN generator. Five-band EEG signals are mapped onto a 4×5 electrode topology, encoded using spatial convolution and channel attention, and divided into temporal patches for self-attention-based modelling.
Fig. 5: Conditional adversarial learning framework. The discriminator distinguishes real EEG–EMG pairs from EEG–generated-EMG pairs and provides distribution-level feedback to the generator.
Dataset
Mean correlation
Std.
WAY-EEG-GAL
0.6496
0.1294
Wrist movement dataset
0.4662
0.0738
TABLE I: Average inter-muscle correlations between the WAY-EEG-GAL dataset and the self-collected wrist movement dataset.
Fig. 6: Inter-channel Pearson correlation matrices of EMG activation in the WAY-EEG-GAL dataset and the self-collected wrist movement dataset. The wrist dataset exhibits weaker and less uniform inter-muscle correlations.
Model
WAY-EEG-GAL
Wrist dataset
Variation (%)
1D-CNN
0.6330±0.1024
0.2463±0.1071
−61.1
CNN-LSTM
0.6803±0.0837
0.2209±0.1194
−67.5
EEGNet
0.7031±0.0753
0.2592±0.0987
−63.1
LSTM
0.6983±0.0814
0.1905±0.0873
−72.7
Transformer
0.7152±0.0829
0.2262±0.1069
−68.4
TABLE II: Cross-task decoding performance of five baseline EEG-to-iEMG models on the WAY-EEG-GAL dataset and the wrist movement dataset. Results are reported as mean ± standard deviation across participants.
Fig. 7: Cross-task decoding performance comparison of five baseline EEG-to-iEMG models on the WAY-EEG-GAL dataset and the self-collected wrist movement dataset. All baseline models exhibit performance degradation when applied to wrist movements.
Model
PCC
nRMSE
1D-CNN
0.2463±0.1071
0.2072±0.0281
CNN-LSTM
0.2209±0.1194
0.2311±0.0387
EEGNet
0.2592±0.0987
0.2026±0.0269
LSTM
0.1905±0.0873
0.2167±0.0346
Transformer
0.2262±0.1069
0.2108±0.0238
ST-Topo GAN
0.4436±0.0981
0.2314±0.0310
TABLE III: EEG-to-EMG decoding performance on the self-collected wrist movement dataset. Results are reported as mean ± standard deviation across participants. The best result for each metric is shown in bold.
Fig. 8: PCC comparison of ST-Topo GAN and representative baseline models on the self-collected wrist movement dataset. Each result represents the average performance across twelve EMG channels for each participant.
Fig. 9: Representative twelve-channel wrist muscle activation reconstruction obtained using ST-Topo GAN. Ground-truth iEMG targets and predicted activation sequences are shown over consecutive testing windows.
Model
PCC
Variation (%)
NoBands
0.4510±0.0802
-17.7
NoTopo
0.4701±0.0356
-14.2
NoAttention
0.5093±0.0347
-7.1
NoGAN
0.5120±0.0266
-6.6
ST-Topo GAN
0.5481±0.0905
–
TABLE IV: Ablation study on three representative participants from the self-collected wrist movement dataset. PCC reduction is calculated relative to the complete ST-Topo GAN.