University of Otago logo. Computer and Information Science Seminars

Seminar Homepage

Speaker:

David (Yi) Wang, Department of Computer Science

Title:

EEG-based Anxious Personality Prediction and Potential Biomarker Visualization using Convolutional Neural Networks

Location:

Owheo G34 - 1:00 pm, Friday 20th September

Abstract:

In recent years, it has been increasingly acknowledged that anxiety disorders, which seriously affect people's quality of life, are a leading source of mental illness. Clinical anxiety disorders are hard to treat since the diagnostic method is questionnaire-based (symptom-based), which means it cannot distinguish specific biological causes of specific disorders. Since non-invasive low-cost scalp electroencephalogram (EEG) recordings can assess anxiety-related brain activity, it is a possible medium for studying the biological causes of anxiety. However, manual searching of these features is laborious, time-consuming, and error-prone. It is essential to design an automatic EEG feature extraction scheme to identify potential anxiety biomarkers.

In this study, two EEG-based Convolutional Neural Network (CNN) architectures have been proposed to predict anxious personality and visualize potential anxiety biomarkers. Intuitively, several questions have been asked.

  1. How to design a psychology theory-based CNN architecture to automatically explore potential features and predict anxious personality?
  2. How to combine the brain topology information into the EEG data structure and design a more general CNN architecture to extract hierarchical temporal-spatial features?
  3. If the proposed models perform well, how to open the 'Blackbox' of the Deep Learning models and visualize decision-making features?
  4. If the models do not perform as expected, how to solve the pervasive (but often neglected) EEG task debugging dilemma, i.e., the Triple-Blind Problem?

Based on these questions, firstly, we will explore a two-dimensional Conflict-focused CNN (2-D CNN) followed by a generalized three-dimensional CNN (3-D CNN). Then, we will open the 'Blackbox' of the models and find out the decision-making components in the input space. Moreover, we will rethink the triple-blind debugging situation and discuss the Validation-Application-Exploration solution.

 

 

Last modified:

This page is maintained by the seminar list administrator.