Biomedical Engineering Reference
In-Depth Information
FIgURE 4.19:
A comparison of misadjustments between the online neuronal channel selection and the
NLMS: (a) 3D data, (b) 2D data.
from neural activity for samples in SEG2 better than the first filter. The statistical test showed that
the null hypothesis was rejected for both 3D and 2D data sets (
P
< 10
−12
). These results imply that
we may need to rebuild the linear filter during the session to track the change of the relationship
between neural activity and kinematics.
Finally, we investigated the misadjustment of two linear adaptive systems, one of which was
adapted using the NLMS and the online subset selection algorithm, whereas the other was adapted
using only the NLMS. For the 3D data, we computed the average misadjustment for each movement
segment. For the 2D data in which the continuous movement was recorded, we arbitrarily divided the
data into 10-sec segments and computed the average misadjustment for each segment. Figure
4.19
shows the results of these average misadjustments for each data set. In this figure, the linear adaptive
system with the online subset selection exhibits smaller misadjustments than the normal linear adap-
tive system for almost every segment. This result is consistent with a previous demonstration of the
superior tracking performance of the online subset selection algorithm, as reported in [
33
].
4.4 SUMMaRy
A reduction in the number of free parameters without affecting performance leads directly to BMI
systems that require less power, bandwidth, and computational demands. Solving these challenges will
bring us one step closer to real, portable BMIs. Regularization and analysis of the most important
neurons in the model-dependent methods has opened up directions to better understand how neural
activity can be effectively used for BMI design. Since the work of Wessberg et al, it is known that the
