Fig. 4

BCI performance and topographic maps. a Evolution over training runs of the decoder accuracy (green, % of correctly classified samples) and rejection (red, % of samples whose prediction was discarded due to low confidence). Their corresponding linear fits and Pearson correlation coefficients (significance tested with Student t test distribution) were evaluated for the two years (2019, 2020) separately. Vertical thin lines indicate the date of each training session, while vertical thick black lines represent the dates of decoder update. The break of 1 year is marked by a vertical red line. (b-c) Boxplots of decoder accuracy b and rejection c in the first and last 15 runs of 2019 and 2020 training periods. The box edges signify the 75th (top) and 25th (bottom) percentiles and the horizontal line the median of the corresponding distribution. The whiskers extend to the largest and smallest nonoutlier values. Single-run values are marked with filled circles. Statistically significant differences are shown with Tukey-Kramer post-hoc tests. d Topographic maps of discriminancy per training month on the 14 EEG channel locations over the sensorimotor cortex. Bright color indicates high discriminancy between Both Hands and Both Feet MI tasks. The discriminancy of each channel is quantified as the Fisher score of the EEG signal’s power spectral density distributions for the two mental classes in the \(\beta\)-band (16–26 Hz) within each run. Each map illustrates local Fisher scores (with interchannel interpolation) averaged over all runs within the month