The role of low-frequency activity in BCIs motor decoding: A comparison between invasive (ECoG) and non-invasive (EEG) brain recordings during repetitive finger movements

Brain-Computer Interfaces (BCIs) enable direct communication between the brain and external devices by interpreting neural signals. BCIs can utilize both invasive signals, such as electrocorticogram (ECoG), and non-invasive signals, like electroencephalogram (EEG). This work aims to compare EEG and ECoG signals during repeated finger flexion-extension. First, it examines CorticoKinematic Coherence (CKC) at low frequencies, analyzing Low Motor Potentials (LMPs). Then, it assesses neural decoding for BCIs by predicting finger trajectories, comparing LMPs and higher frequency bands, as well as a basic (Multiple Linear Regressor, MLR) and a more complex model (Temporal Convolutional Network, TCN). ECoG data comes from the publicly available Stanford ECoG dataset, which includes recordings from nine subjects performing repeated self-paced, single-finger flexion-extension. For EEG data, an experiment was designed, recording signals during repeated finger flexion-extension at 1 Hz and 3 Hz, with movement frequency guided by a shrinking circle. Six subjects’ EEG signals were recorded using a 64-electrodes cap, while finger movements through a sensor glove. ECoG data were provided already preprocessed, with only an additional Common Average Reference (CAR) applied. For the EEG data, preprocessing included removing noisy intervals, applying Notch filters, bandpass filtering, removing eye movement artifacts, and re-referencing with CAR. CKC was then calculated for both datasets. ECoG showed an average coherence of 0.45 ± 0.2, consistent across fingers but varying between subjects, ranging from 0.15 (Subject 5) to over 0.6 (Subjects 2, 3, and 6). EEG exhibited lower coherence, with a mean of 0.02 ± 0.01 across subjects, and 0.03 for Subject 2. Additionally, while CKC frequency trends and topographic distributions for ECoG matched expectations, with EEG, it happened only in a few instances. These results suggest that achieving optimal frequency stability in repeated movements is crucial for analyzing CKC with EEG, while this issue is less impactful with ECoG, highlighting the need for high-quality signals. For the decoding, signals were lowpass filtered at 4 Hz for LMPs, 40 Hz for full-band EEG, and 200 Hz for full-band ECoG. Finger trajectories were predicted using a 0.5 second-brain signal time window preceding the corresponding time instant. Results are reported for the best-performing subject for conciseness. As regards ECoG, the best correlation values between original and predicted trajectories were obtained with the LMPs (0.56 to 0.78 across fingers), with no significative difference using the MLR or TCN. Using the full band, MLR exhibited correlations of 0.44-0.68, while TCN between 0.54 and 0.73. EEG, as expected from the coherence analysis, showed generally low decoding performances with LMPs, with MLR correlations ranging from 0.04 to 0.07 across fingers. TCN did not provide a notable improvement (0.05-0.10). However, some discrete results have been reached by applying TCN to full band data, with correlations between 0.08 and 0.31, while MLR failed. Results suggest that LMPs provide valuable decoding information, with even a simple MLR performing well when the LMPs are clear. If they are unclear, TCN can be useful using instead raw signals with higher frequencies. Future works could focus on ensuring consistent movement frequency during repeated finger movements when analyzing CKC. Moreover, for both signals there seems to be a positive correlation between CKC and decoding performance.