Kernel-Based Relevance Analysis with Enhanced Interpretability for Detection of Brain Activity Patterns

Figures

• 

  Figure 1. 2D representation (top and bottom rows) and topoplots (middle row) of the performed relevance analysis on motor imagery task. The plot at the top shows the marginal relevance (mean and standard deviation) per channel. The topoplots are computed by averaging the input feature relevance values in ϱ regarding the EEG montage (channels). The right-side plot shows the marginal relevance averaged for all considered principles of feature extraction.

• 

  Figure 2. Motor Imagery discrimination accuracy performed by the feature selection strategy. The achieved average accuracy of classification is computed using a nested cross-validation approach, adding one by one the features ranked. VRA, left column; EKRA-S, middle column; DR, right column. Results are displayed concerning each studied subject.

• 

  Figure 3. Contribution of the selected feature set to the Motor Imagery discrimination performance. VRA, left column; EKRA-S, right column Feature extraction principle (PSD, Hjorth, CWT, and DWT parameters) - top row, Brain area (SP and MF) - bottom row. Results are displayed concerning each studied subject.

• 

  Figure 4. Performed accuracy for epileptic seizure detection using each compared approach of feature selection. The achieved average accuracy of classification is computed using a nested cross-validation approach, adding one by one the features ranked. VRA (continuous lines), EKRA–S (dashed lines). The blue and green colors hold for the 2C and the 3C problem, respectively.

• 

  Figure 5. Relevant rhythms regarding the seizure detection tasks. DR - left column. The number of selected features is shown for each provided classification problem (2C and 3C), in blue and red for VRA and EKRA, respectively. VRA-based rhythms selection - middle column, EKRA-S-based rhythms selection - right column. The percentage of selected rhythms are shown in colors for the 2C and the 3C problems.