Date of Award

2025

Document Type

Thesis

Degree Name

MS in Science

Department

Division of Computer Science, Mathematics and Science

First Advisor

Christoforos Christoforou

Abstract

Analysis of electroencephalogram (EEG) recordings in children with dyslexia has been commonly used to explore the neural mechanisms underlying reading disorders. Yet, challenges such as low signal-to-noise ratio (SNR), high inter-subject and inter-trial variability, and the inherently multivariate nature of EEG signals hinder the isolation of neural components elicited during reading diagnostic tests. To mitigate these challenges, the neural-congruency analysis framework was recently proposed, leveraging traditional machine learning optimization methods to incorporate domain knowledge about the congruency of neural responses across participants (i.e., consistent neural responses among proficient readers). However, the application of deep learning techniques, specifically contrastive learning, remains underexplored in analyzing EEG recordings from dyslexic individuals. Motivated by the success of the neural-congruency analysis framework, this study explores the use of a novel deep neural network designed to enforce the neural-congruency constraints and isolate informative neural components in EEG signals. Particularly, our approach integrates spatial, frequency, and temporal convolutions and vectorizations, trained using contrastive learning to extract meaningful neural patterns and identify similarity across different participants’ EEG data. Moreover, the model utilizes a predictive neural network architecture to distinguish neural responses between participants’ grouping. We evaluate our method using synthetically generated EEG data simulating two participant groups (dyslexic and control) performing a shared cognitive task. The data are synthesized across a range of SNR levels from -37 dB to -7 dB to test the model’s robustness under varying noise conditions. Results demonstrate that our network successfully identifies discriminative neural components even at SNR levels as low as –25 dB—substantially noisier than typical real-world EEG recordings—suggesting strong potential for real EEG data applicability. In future work, we plan to apply our methods to three EEG datasets involving dyslexic children performing Rapid Automatized Naming (RAN) and Phonological Processing (PA) tasks. This study represents a significant step toward applying advanced machine learning techniques to complex neural data, offering a novel tool for educational and clinical research on reading difficulties.

Download

Share

COinS