• brain-machine interface
• decoding neural signals
• local field potential
• wavelet scattering transform

When it comes to helping patients with movement-related disorders (e.g. spinal cord injuries, stroke), neural engineering applications such as brain-machine (computer) interfaces (BMIs) hold great promise. One of the key ingredients to building a successful BMIs lies in the ability to read specific user instructions from their brain. Exploiting this information using neural signals is crucial to send commands to an output device (such as a robotic arm) to perform the desired action. The issue to understand how we can read the desired action from the neural activity is called neural decoding. Nowadays, implementing sophisticated decoding algorithms for proper detection and translation of neural signals into suitable instructions is a central topic for neural engineering applications, for example, the development of fully functional BMIs.

Here, we implemented an advanced decoding algorithm to decode distinct phases of movement in an instructed-delay reach-to-grasp task performed by a non-human primate. Specifically, we were interested to decode four distinct epochs, i.e. Free, Obj-Vis, Delay, Grasp, in two different conditions (light and dark conditions).

Briefly, the Free period is related to a rest condition, Obj-Vis and Delay to movement preparation and Grasp to movement execution. In the light condition, the animal can see the object to grasp whereas in the dark condition the object is visible only for a short period.

Neural signals were acquired using a multielectrode array implanted in the primary motor cortex (M1) of a Macaca fascicularis. In particular, as a signal for our decoding task, we used the Local Field Potential (LFP) that represents the low frequency (=<300 Hz) part of the whole extracellular potential. An innovative decoding algorithm based on wavelet scattering transform, principal component analysis, and support vector machines was proposed to decode the different epochs of the experiment.

The results show reliable decoding of the distinct four epochs of the experiment with an average accuracy of 82% and 83% in the light and dark conditions, respectively. We also found no statistically significant difference between the two conditions, indicating that our decoding algorithm is robust to variations of the task. To the best of our knowledge, no other study has yet employed wavelet scattering transform to extract movement-related neural patterns from the LFPs in an instructed-delay reach-to-grasp task. Our results could represent an important step toward the development of BMI that relies on the detection of movement actions.