Figure 7. Algorithmic contact prediction vs beta activity as single feature.

This illustrates the percentage change of the prediction performance of the presented algorithmic approach relative to the use of beta activity as single feature. The relative change of the prediction performance has been derived as change in the area under the curve for testing only one and up to three contacts out of six segmented contacts (a) as well as for testing one and up to four out of all eight contacts (b). The maximum algorithmic prediction significantly outperforms the use of beta activity for the first-choice contact (1/6: CE [117%] and TW [30%]; 1/8: CE [99%], TW [57%], and ST [25%]) and for half of the contacts (3/6: CE [46%], TW [15%], and ST [20%]; 4/8: CE [32%], TW [17%], and ST [21%]). The average algorithmic prediction outperforms the use of beta activity for first-choice contact (1/6: CE [63%]; 1/8: CE [38%] and TW [14%]), and for half of the contacts (3/6: CE [20%]; 4/8: CE [5%]). The minimum algorithmic prediction outperforms beta activity for the first-choice contact (1/6 CE [17%]). In the remaining iterations, the algorithmic approach does not show an advantage over beta activity used as a single feature. Improvement is illustrated as % median improvement. Statistical comparison was performed as a one-sampled t-test (false discovery rate corrected) for all iterations to test whether the prediction is significantly above or below zero (zero corresponds to the prediction obtained by low beta activity). **p < 0.01; ***p < 0.001. Detailed statistics in the Supplementary Material. [Color figure can be viewed at www.neuromodulationjournal.org]