Lahiru Wimalasena
BME PhD Defense Presentation
Date: 2023-11-07
Time: 10:00 AM - 12:00 PM
Location / Meeting Link: HSRB II Room N600 / https://emory.zoom.us/j/91715756801
Committee Members:
Chethan Pandarinath, PhD (Advisor); Nicholas Au Yong, MD, PhD (Co-advisor); Samuel Sober, PhD; Lena Ting, PhD, FAIMBE; Gordon Berman, PhD; Lee Miller, PhD
Title: Methods and analyses to uncover the muscle pattern-generating mechanisms of the spinal cord and motor cortex using deep learning-based dynamical systems models
Abstract:
The motor nervous system can flexibly generate a wide range of voluntary movements by coordinating the time-varying activity of muscles with precision on the order of milliseconds. Understanding the natural muscle pattern-generating functions of the motor cortex and spinal cord is critical for the development of technologies like brain-machine interfaces that aim to restore motor function through the stimulation of muscles. The objectives of this thesis were to perform high-fidelity analyses of the neural population activity in the spinal cord and motor cortex to understand their roles in generating patterns of muscle activations. The first aim of this research was to develop methods to generate high-fidelity muscle activation estimates from EMG recordings. We adapted a large-scale deep learning-based dynamical systems model optimization framework for cortical spiking activity (AutoLFADS) and demonstrated its broad application to de-noise multi-muscle EMG recordings. The second aim of this research was to investigate the activity of spinal interneuron populations to understand their role in locomotor pattern generation. In chapter 3, we pioneered application of AutoLFADS to spinal interneuron and muscle recordings to uncover precise relationships. The third aim of this research was to study the relationship between motor cortical populations and muscle activations. In chapter 4, we studied the extent to which linear readouts could predict muscle activations from M1 during a complex reach-to-grasp task. Finally, in chapter 5, we developed a platform to perform manifold alignment of M1 population activity that led to stable prediction of muscle activations over 95 days.