Mapping broadband electrocorticographic recordings to two-dimensional hand trajectories in humans Motor control features.

Brain-machine interfaces (BMIs) aim to translate the motor intent of locked-in patients into neuroprosthetic control commands. Electrocorticographical (ECoG) signals provide promising neural inputs to BMIs as shown in recent studies. In this paper, we utilize a broadband spectrum above the fast gamma ranges and systematically study the role of spectral resolution, in which the broadband is partitioned, on the reconstruction of the patients' hand trajectories. Traditionally, the power of ECoG rhythms (<200-300 Hz) has been computed in short duration bins and instantaneously and linearly mapped to cursor trajectories. Neither time embedding, nor nonlinear mappings have been previously implemented in ECoG neuroprosthesis. Herein, mapping of neural modulations to goal-oriented motor behavior is achieved via linear adaptive filters with embedded memory depths and as a novelty through echo state networks (ESNs), which provide nonlinear mappings without compromising training complexity or increasing the number of model parameters, with up to 85% correlation. Reconstructed hand trajectories are analyzed through spatial, spectral and temporal sensitivities. The superiority of nonlinear mappings in the cases of low spectral resolution and abundance of interictal activity is discussed.