Fig. 8

Summary results of the online experiment. a) Median and interquartile range (25–75 %) of the normalized mean square error between trapezoidal reference trajectory and control signals during active phases, and the mean amplitude of the false activations during silent phases across control methods for the eight finger movements. Overall, the online performance of the two machine learning methods was similar and better than the THR. b) Median and interquartile range (25–75 %) of the summary results shown in A. c Radar graph with the results for the accuracy of tracking (nMSE) for each subject. Each spoke represents one of the target finger movements and the length of the spokes is normalized to the maximum value of nMSE. The radar graphs for THR are characterized with spikes indicating larger errors, but the spikes are limited to some fingers (one, two typically), while for the other fingers the performance is actually comparable to that of the machine learning methods. *p < 0.01. The results of the distribution analysis are reported as circles ° p < 0.01