Dorsal Column Nuclei Neural Signal Features Permit Robust Machine-Learning of Natural Tactile- and Proprioception-Dominated Stimuli

Abstract

Read online

Neural prostheses enable users to effect movement through a variety of actuators by translating brain signals into movement control signals. However, to achieve more natural limb movements from these devices, the restoration of somatosensory feedback is required. We used feature-learnability, a machine-learning approach, to assess signal features for their capacity to enhance decoding performance of neural signals evoked by natural tactile and proprioceptive somatosensory stimuli, recorded from the surface of the dorsal column nuclei (DCN) in urethane-anesthetized rats. The highest performing individual feature, spike amplitude, classified somatosensory DCN signals with 70% accuracy. The highest accuracy achieved was 87% using 13 features that were extracted from both high and low-frequency (LF) bands of DCN signals. In general, high-frequency (HF) features contained the most information about peripheral somatosensory events, but when features were acquired from short time-windows, classification accuracy was significantly improved by adding LF features to the feature set. We found that proprioception-dominated stimuli generalize across animals better than tactile-dominated stimuli, and we demonstrate how information that signal features contribute to neural decoding changes over the time-course of dynamic somatosensory events. These findings may inform the biomimetic design of artificial stimuli that can activate the DCN to substitute somatosensory feedback. Although, we investigated somatosensory structures, the feature set we investigated may also prove useful for decoding other (e.g., motor) neural signals.

Keywords

• feature learnability
• neural prosthesis
• supervised back-propagation artificial neural network
• brain-machine interface
• cuneate
• gracile