Deep spatio-temporal features optimised fusion with coordinate attention mechanism for EEG lower limb pre-movement intention decoding

Background and Objective

Decoding pre-movement intention is crucial in developing a brain-computer interface (BCI) for neuro-rehabilitation robotic systems. However, the weak amplitude and non-smooth characteristics of EEG signals lead to the inability of existing methods to achieve the accuracy for proper applications. This study proposed a novel pre-movement intention decoding network framework to improve accuracy by extracting and optimizing the deep spatio-temporal features of EEG signals.

Methods

A deep spatio-temporal neural network structure was constructed based on the brain intention generation mechanism and its movement expression. The collected multi-channel EEG data were reorganized into brain topographic distributions, after the initial extraction of the features and optimization using the coordinate attention mechanism, a 3-layer dense block with two bi-directional gated recirculation units was designed to effectively extract the deep spatial and temporal features, further decoding the pre-movement intention efficiently.

Results

The experimental results showed an average accuracy of 95.51 ± 1.79 % for healthy subjects and 90.48 ± 2.90 % for stroke survivors in decoding pre-movement intention. All evaluation indexes are excellent. Pseudo-online testing showed the average TPR was 95.45 ± 3.80 % and 90.71 ± 7.77 % for healthy subjects and stroke survivors, respectively, and the latency was −1965 ± 48 ms and −1974 ± 36 ms. The results of the ablation and comparative analysis showed that the proposed framework is justified and its decoding capability outperforms other state-of-the-art algorithms.

Conclusion

The method proposed in this study has high decoding accuracy and good online performance in pre-movement intention decoding based on EEG signals, which lays the foundation for further neuro-rehabilitation robotic systems.