Abstract:Motor decoding at the level of individual finger movements is critical for high-performance brain-machine interface (BMI) applications. In this work, we propose to exploit the temporal dynamics of the multi-channel electrocorticography (ECoG) signal from human subjects and modern machine learning algorithms to improve the finger-level movement classification accuracy. Using a decision tree ensemble as the classifier and the temporally-concatenated features of ECoG as input, we achieved an average classification accuracy of 71.3%±7.1% on 3 subjects, 6.3% better than the state-of-the-art approach based on conditional random fields (CRF) on the same dataset. Our proposed method could enable a high-performance and minimally invasive cortical BMI for paralyzed patients.

Enhanced Classification of Individual Finger Movements with ECoG.