Machine Learning to Characterize Motor Patterns and Restore Walking after Neural Injury

  • Author / Creator
    Dalrymple, Ashley
  • Walking is a locomotor task that integrates information from all over the nervous system. The lumbosacral spinal cord houses neural networks that contribute to locomotion. These networks dominate locomotor activity during development and may provide suitable targets for restoring function after injury. Motor activity of the developing spinal cord under electrical or pharmacological activation has been extensively used to study locomotor networks, leading to various models of the central pattern generator. Spontaneous activity of the developing spinal cord, which represents the locomotor networks at rest, may also contribute to understanding of the development of these networks. In the postnatal rodent spinal cord, it is characterized by stochastic and complex patterns of activity, which correspond to kicking-like movements. Spontaneous activity has been challenging to characterize as there are no current analysis methods. I developed a software tool to characterize and classify episodes of spontaneous activity from developing spinal networks of the neonatal mouse using supervised machine learning. I tested the software’s ability to detect changes in activity by increasing network excitability with KCl. Supervised machine learning-based classification revealed global and class-specific changes after increasing excitability. This software will add to the toolbox of methods used to study developing locomotor networks under varying conditions.After a SCI, spinal neural networks and their connections to the leg muscles remain intact. An incomplete SCI causes partial paralysis. To restore walking after an incomplete SCI, residual function needs to be augmented. Intraspinal microstimulation (ISMS) entails implanting electrodes into the ventral horn of the lumbosacral enlargement to activate the muscles of the legs. I used ISMS to restore walking in a model of hemisection SCI, which affects one hind-limb. Anaesthetized cats with an intact cord were implanted with ISMS unilaterally and the voluntary movements of one hind-limb were mimicked by a person moving that limb. Feedback from external sensors on the person-moved limb, representing residual function, was used to move the other limb to the opposite phase of the gait cycle using ISMS. The first demonstration of augmenting remaining function was performed in cats on a split-belt treadmill. The belt ipsilateral to the ISMS-controlled limb remained stationary, while the other turned at varying speeds. Sensors measuring residual function of the person-moved limb were used to anticipate changes of the walking phases, and triggered ISMS to move the other limb to the opposite phase. At faster speeds of stepping, the feedback-initiated transitions were insufficient causing a loss of weight-bearing. Four different supervised machine learning methods were used to predict the step period of the person-moved limb. If the prediction indicated a faster step, the control strategy changed to feed-forward, using the predicted value to determine the time spent in each phase of the cycle. Three of the four prediction methods resulted in improved weight-bearing, and maintained alternation at varying speeds. This control strategy augmented remaining function, while allowing the user to step at a self-selected speed through automatic adaptation using supervised machine learning.For more personalized control of walking, other machine learning methods were employed. Commercially available walking systems use the same open-loop control strategy for each user, forcing them to accommodate to the control system. Feedback from external sensors can improve control strategies; however, with a large burden of tuning, as each person walks differently from others as well as among themselves. I propose that machine learning can reduce the burden of tuning and demonstrate this using ISMS in a feline model of hemisection SCI. Reinforcement learning generated predictions for sensors measuring residual function during walking. Thresholds on the predictions indicated phases of the walking cycle and produced fixed responses in the ISMS output to move the other limb to the opposite phase (Pavlovian control). Learning parameters were either initialized to zero or built upon learned predictions from previous walking trials. Predictions were quickly learned and initiated changes between the phases of the walking cycle to produce alternating over-ground walking. Learning was able to adapt to different people walking the limb and between different cats. Furthermore, learning was able to recover from mistakes made during walking. This work demonstrated that Pavlovian control using reinforcement learning can adapt to different subjects walking without the need for retuning. It allowed for personalized walking that augmented remaining function.

  • Subjects / Keywords
  • Graduation date
    Spring 2019
  • Type of Item
  • Degree
    Doctor of Philosophy
  • DOI
  • License
    Permission is hereby granted to the University of Alberta Libraries to reproduce single copies of this thesis and to lend or sell such copies for private, scholarly or scientific research purposes only. Where the thesis is converted to, or otherwise made available in digital form, the University of Alberta will advise potential users of the thesis of these terms. The author reserves all other publication and other rights in association with the copyright in the thesis and, except as herein before provided, neither the thesis nor any substantial portion thereof may be printed or otherwise reproduced in any material form whatsoever without the author's prior written permission.