Fig. 2: Random forest classifier features and their evolution over training.

a Relationship between individual BCI-control performance and feature weights, obtained by considering the sum of the first 200 features, averaged across the 5 training days (n = 15). b Average of features’ weights for each frequency across all experimental days, with colors indicating the different frequency bands. The highlighted intervals indicate two frequency ranges with a prominent contribution to syllable decoding: 8–16 Hz and 28–70 Hz. c Corresponding topographies of the average weights for these two frequency intervals. d Maps visualizing the average feature weights across participants, for each training day. On each map, channels are represented on the y-axis, and individual frequency values on the x-axis. Scalp topographies below each map display the average weight for each channel, across all frequencies and participants, on each training day. e The bar plot represents statistical results (t values) when testing for changes in frequency contribution with training. Positive and negative values indicate respectively a decrease and increase in contribution throughout the training period. Yellow bars highlight the frequency(ies) for which the test was statistically significant (p value < 0.05, permutation test, n = 15). Scalp topographies show the results of the same statistical analysis performed on the scalp space, separately for two frequency intervals. These were defined based on the analysis in the frequency space, and include respectively the 2–10 Hz (low-frequency interval of interest) and 52–66 Hz (high-frequency interval of interest) range. A decrease in the contribution of a given electrode in the classification is indicated in blue, an increase in red (t values). Black dots highlight electrodes showing a statistically significant effect (p value < 0.05, permutation test, n = 15). f Schematics of the approach used to extract a global index to quantify the change across the 5 days of training. For each feature in the channel x frequency space, the Euclidean distance between two consecutive training days is calculated (as indicated by the red dotted line), and then the index is obtained by averaging the resulting 4 distances. g Euclidean distance index for each individual feature. h Correlation between the global index representing the amount of change both in the frequency and spatial domains (computed as the Euclidean distance between the features’ weights from two consecutive days) and the average BCI performance across the 5 training days (n = 15).