Improving Decoding of Electrocorticographic Signals using Deep Learning Restricted; Files Only

Li, Kejun (Spring 2019)

Permanent URL: https://etd.library.emory.edu/concern/etds/bv73c157f?locale=en
Published

Abstract

Brain-machine Interfaces (BMIs) are technologies that aim to assist people who suffer from loss of motor function due to accidents or neurodegenerative diseases, by providing a pathway between brain and external assistive devices. Electrocorticography (ECoG) is one of the recording methodologies that could be used to infer people’s movement intention for BMI systems. ECoG covers large areas of the brain and has a demonstrated safety record. However, its low spatial resolution usually leads to worse accuracy in decoding movement intention than BMIs that use intracortical recordings. In this study, we found that ECoG decoding performance could be improved by modeling neural population dynamics using a deep learning technique: Latent Factor Analysis via Dynamical Systems (LFADS). LFADS attempts to uncover structure from neural population activity that is consistent with a low-dimensional dynamical system. In previous applications to intracortical recording from motor cortex, LFADS improved decoding accuracy by uncovering estimates of neural population dynamics on a single-trial, moment-by-moment basis. However, since LFADS was previously evaluated using intracortical recordings with high spatial resolution, it was unclear whether LFADS is appropriate for de-noising recordings with low spatial resolution, such as ECoG. To test this, we applied LFADS to ECoG recordings from seven human subjects who were being monitored as part of clinical treatment for epilepsy or glioma. Subjects performed a one-finger button-press task as their finger kinematics and kinetics were recorded. We compared our ability to decode behavioral states (i.e. pre-movement, movement or force) before and after application of LFADS. The accuracy of the discrete classifier was improved significantly by the application of LFADS. Our results represent a new avenue toward improving the performance of ECoG-based BMI systems. 

Table of Contents

Terminology and Abbreviations …..………………………………………….…………… 1

Introduction …………………………………………….………………………….……… 2 - 5

Methods …………………………………………….………………………….………… 6 - 14

1.   Subjects and Recording

2.   Experimental Protocol

3.   Spectral Feature Extraction

4.   Spectral Feature Extraction and Power Amplitude (PA) Matrix

5.   LFADS Structure

6.   Hyperparameter Tuning with Population-based Training

7.   Principal Component Analysis

8.   Decoding Analysis

Results …………………………………………….………………………….………… 15 - 17

Figures …………………………………………….………………………….………… 18 - 25

1.   Demonstration of PCA analysis

2.   Decision tree diagram

3.   Experimental protocols and recording ECOG array placement for the subjects.

4.   Scheme for discrete states classifiers

5.   Percent variance explained by the first three principal components from PCA for PA matrix and LFADS-denoised ECoG

6.   Example trial for ECoG and PCA trajectories of PA matrix and LFADS-denoised ECoG

7.   Bagged trees decoding performance

Discussions …………………………………………….………………………….…... 26 - 29

References …………………………………………….………………………….……. 30 - 33

About this Honors Thesis

Rights statement
  • Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.
School
Department
Degree
Submission
Language
  • English
Research field
Keyword
Committee Chair / Thesis Advisor
Committee Members
Last modified No preview

Primary PDF

Supplemental Files