Figure 5
Factors affecting the feasibility of BCI free communication, assessed using QWERTY classification data. (a) Grand mean classification features (Ri) for each cued frequency for the high and low accuracy groups. For each flicker period, the filter-bank CCA produced a classification feature (Ri) representing the correlation between the single-trial EEG and templates at each frequency. Correct classification occurred when Ri was maximal for the template matching the cued frequency. Template frequencies in the alpha range are indicated by the orange bounding boxes (dashed lines). (b) Box plots showing the improvement in classification accuracy using the optimal number of harmonics rather than the fixed first five harmonics that were used in real-time. (c) Simulated classification accuracy by number of harmonics, plotted for the high and low accuracy groups. (d) Simulated classification accuracy by number of training trials, plotted for the high and low accuracy groups. The fitted model is an inverse exponential function. (e) Cross-participant classification. To assess template generalisability, each participant’s single-trial EEG was classified using all other participants’ templates. The leading diagonal represents accuracy when participants were classified using their own templates. The leftmost column (read bottom to top) represents the highest accuracy data classified with progressively lower accuracy templates. Similarly, the bottom row (read left to right) represents the highest accuracy templates used to classify progressively lower accuracy data.
