Classifying Autism Spectrum Disorder via Brain Connectivity and Graph Neural Networks

Frey Pereira, Valesca (2025) Classifying Autism Spectrum Disorder via Brain Connectivity and Graph Neural Networks. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (2MB) | Preview
Preview	PDF (Configuration Manual) Download (1MB) | Preview

Abstract

Autism Spectrum Disorder (ASD) is a neurodevelopmental condition where the one has challenges in social interactions, communication, and behaviours. Current researches show that male are more frequently diagnosed than female, with ratios up to 4:1. This imbalance raises an important concern: are the models adjusted in generalised data effectively classifying ASD in women? While the field has gained advances with the use of Graph Neural Network (GNN) models, the question of weather these models perform equally well considering sexes is largely unexplored. This study investigates whether autistic women exhibit different neural connectivity patterns that require sex-specific diagnostic modelling. To explore this, fMRI data from the ABIDE I dataset were converted into brain graphs using the AAL atlas, combining functional correlations with structural features. The study began by training a Graph Convolutional Networks (GCN) and a Graph Attention Networks (GAT) as baseline for the future experiments. A series of training scenarios were applied to verify the sex differences systematically. Models were trained on men only, on all subjects combined, on women only, and on a balanced subset. Their performance was evaluated through stratified cross-validation and held-out testing. Across all scenarios, accuracy was consistently lower for women. When trained on men subjects only, the precision showed a gap of 20% men compared to women. In the balanced subset, women’s performance improved, likely because they were more equally represented, although the small dataset and a 10-fold split also introduced variability. In the women only trained model, the accuracy improved to 70%. Overall, the results indicate that autistic women might have distinct connectivity shapes and might therefore require diagnostic models that are explicitly sex-aware. This work highlights the need for better female representation in neuroimaging research and suggests that future ASD classification studies should consider sex-specific modelling approaches.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Simiscuka, Anderson UNSPECIFIED
Subjects:	Q Science > QA Mathematics > Electronic computers. Computer science T Technology > T Technology (General) > Information Technology > Electronic computers. Computer science R Medicine > Diseases > Disabilities > Developmental disabilities Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	07 Sep 2026 10:40
Last Modified:	07 Sep 2026 10:40
URI:	https://norma.ncirl.ie/id/eprint/9857

Actions (login required)

View Item