Figure 3. Feature ranking.

Illustrates the output of step one of the two-step contact prediction method, corresponding to the ranked weights of the spectral features as determined by the Lasso regression. The features are ranked in descending order of predictive value for segmented contacts (a) and all contacts (b) and separately for the different clinical parameters (CE, TW, ST). The weights are averaged across the various iterations (Fig. 1c). The two most predictive features for each depicted configuration are the following: the most predictive feature for CE for both (a) and (b) is resting-state fast gamma activity (negative relationship), followed by resting-state low beta activity (positive relationship). For TW, the feature ranking results are similar, with the most predictive features being resting-state gamma activity (negative relationship) in (a) and resting-state fast gamma activity (negative relationship) in panel b. The second most predictive feature for both contact configurations is high beta activity (positive relationship). For the ST, the most predictive feature is movement-state HFO (negative relationship), followed by gamma at rest (negative relationship) in panel a. In panel b, the most predictive feature is low beta ERD (negative relationship), followed by movement-state HFO (negative relationship). [Color figure can be viewed at www.neuromodulationjournal.org]