Figure 2
From: Leveraging neural dynamics to extend functional lifetime of brain-machine interfaces
Offline and online evaluation of decoder hysteresis. (a) An offline simulation with Monkey J, where electrode loss is simulated and offline decode performance (mean correlation in reconstructing hand velocity) is measured. When 90 or more neural electrodes were lost, the HNDF achieved significantly higher offline decode performance than the NDF (denoted by *p < 0.01, Wilcoxon signed-rank test). The inset shows an example of decoded x-velocity (true hand velocity in gray, NDF decoded velocity in blue, and HNDF decoded velocity in red) when 140 electrodes were lost. In this example, the NDF essentially loses a degree-of-freedom of control, being unable to generate velocities in the x-direction, while the HNDF rescues the BMI by recovering the lost degree-of-freedom. (b) Same as (a) but for Monkey L. When 40 or more neural electrodes were lost, the HNDF achieved significantly higher offline decode performance than the NDF (p < 0.01, Wilcoxon signed-rank test). The inset shows an example of decoded x-velocity when 50 electrodes were lost. (c) The performance of the HNDF, NDF, and FIT-KF in closed-loop experiments for Monkey J, with the simulated loss of 110 electrodes. The HNDF performs significantly better (19.4 radial targets per minute) than the NDF (6.0 radial targets per minute) and FIT-KF, which was uncontrollable on 19 out of 20 sessions. Datasets J_2015-05-18, J_2015-05-19, J_2015-05-20, J_2015-05-21 comprising 3,642 NDF trials and 3,903 HNDF trials. (d) Same as (c) but for Monkey L, with the simulated loss of 60 electrodes. The HNDF performs significantly better (13.9 radial targets per minute) than the NDF (8.0 radial targets per minute) and the FIT-KF, which was uncontrollable on 14 out of 16 sessions. Datasets L_2015-05-28, L_2015-05-29, L_2015-06-02, L_2015-06-03, L_2015-06-04 comprising 4,568 NDF trials and 4,646 HNDF trials. (e) Success rate of radial target acquisition for Monkey J. The HNDF acquired targets at far higher success rates (83%) than the NDF (48%). (f) Same as (e) but for Monkey L. The HNDF acquired targets at far higher success rates (76%) than the NDF (52%). (g) Target success rate to each center-out target in the workspace averaged over all sessions for Monkey J. There was a deficiency in the NDF to the upper and left parts of the workspace. (h) Same as (g) but for the Monkey L. There was a deficiency in the NDF along the x-axis.
