Figure 5. LFP-based contact prediction for all contacts.

a. Illustrates the output of the second step of the contact prediction method, corresponding to the probability of identifying the best stimulation contact for the three clinical parameters (CE,TW, ST) out of the eight contacts across. Within each subplot, the maximum, average, and minimum prediction accuracies evaluated on the hold-outset of nine hemispheres are illustrated as mean value, averaged across the multiple iteration of the prediction pipeline (Fig. 1). The dashed black lines illustrate the prediction by chance (conventional gold standard test strategy), where the probability of identifying the most efficient stimulation contact increases by 0.125 after each contact tested. By applying the LFP-based contact prediction strategy and considering half of the stimulation contacts, the probability of identifying the best stimulation contacts can reach a maximum as follows: CE: 83%, TW: 86%, and ST: 87%. These results are derived as overall output of the prediction algorithm, after combining the five best ranked features, without differentiating whether the best selected features or the stepwise combination of features more strongly predicts the optimal stimulation site. b. Illustrates the percentage change in contact prediction, considering half of the stimulation contacts (4/ 8), following the combination of multiple features relative to the use of the single highest ranked feature for the three clinical parameters (CE, TW, ST). For CE, adding additional features leads to a reduction of the prediction performance (Friedmann test: x² (3) = 53.94, p ≤ 0.001; significant one-sampled t-test for number of two, three, four, five features). For TW (Friedmann test: x (3) = 50.25, p ≤ 0.001; significant one-sampled t-test for number of two, three, four, five features), the prediction performance can be increased by combining multiple features. For ST, the use of more than one feature increases the prediction performance (significant one-sampled t-test for number of two, three, four, five features); however, there is no further benefit in using more than two features for the prediction (Friedmann test: x² (3) = 5.19, p = 0.159). p Values were false discovery rate corrected. *p < 0.05; **p < 0.01; ***p < 0.001. Detailed statistics in the Supplementary Material. [Color figure can be viewed at www.neuromodulationjournal.org]