Peripheral Nerve Activation Evokes Machine-Learnable Signals in the Dorsal Column Nuclei

Abstract

The brainstem dorsal column nuclei (DCN) are essential to inform the brain of tactile and proprioceptive events experienced by the body. However, little is known about how ascending somatosensory information is represented in the DCN. Our objective was to investigate the usefulness of high-frequency (HF) and low-frequency (LF) DCN signal features (SFs) in predicting the nerve from which signals were evoked. We also aimed to explore the robustness of DCN SFs and map their relative information content across the brainstem surface. DCN surface potentials were recorded from urethane-anesthetized Wistar rats during sural and peroneal nerve electrical stimulation. Five salient SFs were extracted from each recording electrode of a seven-electrode array. We used a machine learning approach to quantify and rank information content contained within DCN surface-potential signals following peripheral nerve activation. Machine-learning of SF and electrode position combinations was quantified to determine a hierarchy of information importance for resolving the peripheral origin of nerve activation. A supervised back-propagation artificial neural network (ANN) could predict the nerve from which a response was evoked with up to 96.8 ± 0.8% accuracy. Guided by feature-learnability , we maintained high prediction accuracy after reducing ANN algorithm inputs from 35 (5 SFs from 7 electrodes) to 6 (4 SFs from one electrode and 2 SFs from a second electrode). When the number of input features were reduced, the best performing input combinations included HF and LF features. Feature-learnability also revealed that signals recorded from the same midline electrode can be accurately classified when evoked from bilateral nerve pairs, suggesting DCN surface activity asymmetry. Here we demonstrate a novel method for mapping the information content of signal patterns across the DCN surface and show that DCN SFs are robust across a population. Finally, we also show that the DCN is functionally asymmetrically organized, which challenges our current understanding of somatotopic symmetry across the midline at sub-cortical levels.

Novel neural signal features permit robust machine-learning of natural tactile- and proprioception-dominated dorsal column nuclei signals

“…Abbreviations: DCN, dorsal column nuclei; e1, electrode 1; e2, electrode 2; e3, electrode 3; e4, electrode 4; e5, electrode 5; e6, electrode 6; e7, electrode, 7; SEM, standard error of the mean. (Loutit et al, 2019) are shown. Individual features were divided into four categories (colour coded): four HF (A-D); five LF (E-I); eight HF PSD (J); and five LF PSD (K).…”

“…Feature-learnability was derived from dorsal column nuclei signal features extracted from 1000 ms windows starting at the onset of each stimulus. Features are ordered in featurelearnability rank order (Loutit et al, 2019) from highest to lowest feature-learnability from left to right. Colours indicate the feature category that individual features belong (as described in Fig 2).…”

“…We previously demonstrated that combinations of LF and HF signal features improve machinelearning outcomes (Loutit et al, 2019). We therefore sought to determine a combination of input features that produced the highest feature-learnability, and to establish a benchmark for subsequent comparisons.…”

“…The dorsal column nuclei (DCN) are a potential neural prosthetic target for restoring somatosensory feedback, and have recently begun to receive attention for this purpose (Richardson et al, 2015Sritharan et al, 2016;Loutit et al, 2017;Suresh et al, 2017;Loutit et al, 2019). Determining the relevance of DCN neural signal features to natural somatosensory stimuli from which they were evoked, provides insight into the neurophysiology and functional organisation of the DCN, which may inform potential neural prosthetic research in this region.…”

“…Critically, the neural population, time window sampled, and signal features used, must represent information that is highly relevant to the desired motor behaviour or presented stimulus (Wessberg et al, 2000;Carmena et al, 2003;Homer et al, 2013;Lebedev & Nicolelis, 2017). We previously devised a metric for quantifying the relevance of neural signal features to the stimulus from which they were evoked, which we termed feature-learnability (Loutit et al, 2019). This approach uses a simple feedforward back-propagation supervised artificial neural network (ANN), which has several advantages over the use of state-of-the-art deep neural network (DNN) architectures (Pandarinath et al, 2018): An ANN enables the quantification of information content from signal features of interest, that, for example, may be selected on the basis of: i) their neurophysiological relevance, ii) their capacity to mimic or inform electrical stimulation in sensory applications, and/or iii) are commonly used for decoding neural signals.…”