Towards real time efficient and robust ECoG decoding for mobile brain–computer interface

• Published in: Journal of neural engineering Vol. 22; no. 4; pp. 46012 - 46028
• Main Authors: Lin, Zhanhui, Jiang, Xinyu, Dai, Chenyun, Jia, Fumin
• Format: Journal Article
• Language: English
• Published: England IOP Publishing 01.08.2025
• Subjects: Adult; Algorithms; BCI; Brain - physiology; Brain-Computer Interfaces; brain–computer interface; Computer Systems; ECoG; electrocorticography; Electrocorticography - methods; Female; Humans; iEEG; Male; BCI; brain–computer interface; electrocorticography; ECoG; iEEG
• ISSN: 1741-2560; 1741-2552; 1741-2552
• DOI: 10.1088/1741-2552/ade917

Objective . Decoding locomotion-related brain activities from electrocorticographic (ECoG) signals is essential in brain–computer interfaces (BCIs). Most previous ECoG decoders are computationally demanding and sensitive to noises/outliers. Mobile and robust BCIs are particularly important for physically disabled patients to restore motor ability in outdoor scenarios, where the processing pipeline should be implemented efficiently using constrained computation resources. In this work, we aim to explore the optimal pipeline for mobile BCI decoding. Approach . We comprehensively evaluated the trade-off between the decoding precision, computational efficiency and robustness of diverse decoding algorithms on a combined ECoG dataset of 12 subjects conducting individual finger movement, including partial-least-square and their N-way variants, Bayesian ridge regression, least absolute shrinkage and selection operator, support vector regression, neural networks (NNs) with diverse architectures, and random forest (RF). We further explored the feature optimization technique for selected models by using their inherent model explainability. We also compared the decoding performance of updatable algorithms when the data is split into multiple batches and used sequentially. Main results . The RF model, not valued by previous studies, can achieve the best trade-off between precision and efficiency, achieving an average Pearson’s correlation coefficient ( r ) of 0.466 with only 0.5 K floating-point operations per second (FLOPs) per inference and a model size of 900KiB. We also demonstrate the inherent superior robustness of RF model on corrupted ECoG electrodes, with a > 2 × decoding precision on noisy signals compared with all state-of-the-art deep NNs. The optimized RF processing pipeline was deployed on a STM32-based embedded platform with only a 15.2 ms computation delay. Significance . In this study, we systematically explored the performance and efficiency of ECoG decoding algorithms in finger movement analysis. The proposed decoding pipeline is implemented on a compact embedded platform to achieve low-latency, power-efficient real-time decoding. This research accelerates the translation of mobile BCI into real-life practices.