A design and implementation framework for unsupervised high-resolution recursive filters in neuromotor prosthesis applications

Neuromotor prostheses have the potential of restoring movement ability in patients with severe motor dysfuncion. In cortically-controlled neuromotor prostheses, the design of neural decoders for motor impaired patients requires initialization using a concurrently measured set of neural and motor imagery or observation data. In addition, the decoder implementation poses a scalability challenge with an increasing number of decoded neurons. Consequently, most neural decoder implementations resort to sub-sampling the neural firing rates, which results in noisy decoded outputs. In this work, we propose a new decoder design and implementation framework in which (i) the decoder initialization is unsupervised, (ii) the decoder is implemented using computationally-inexpensive recursive filters that can operate at high-resolution sampling of the neural data thereby minimizing the delay introduced in the system, and (iii) the decoder gives a smooth real-time control signal expressed by the span of neural data projections onto a low-dimensional latent space that possesses desirable features for the control task.