Rupp, S. M. (2026). Autism Spectrum Disorder Classification: As Good As It Gets [Diploma Thesis, Technische Universität Wien]. reposiTUm. https://doi.org/10.34726/hss.2026.146609
Autism Spectrum Disorder (ASD) is associated with heterogeneous modifications in functional brain connectivity that are difficult to characterise using conventional diagnostic approaches. Resting State-Functional Magnetic Resonance Imaging (rs-fMRI) and Machine Learning (ML) provide tools for investigating these alterations; however, the analysis of multi-site neuroimaging data is complicated by substantial acquisition site-heterogeneity and limited cross-site generalisation. This thesis investigates the robustness of ML-based ASD classification in the multi-site Autism Brain Imaging Data Exchange (ABIDE) 1 dataset and examines whether graph-based representations can provide additional insight into the heterogeneous organisation of functional connectivity. Initially, a systematic preprocessing and evaluation framework is established using Stratified Group K-Fold (SGKF) and Leave One Site Out (LOSO). Frequency-domain analysis, confound regression, and Covariate adjustment method Batch (ComBat) harmonisation are evaluated for their ability to reduce acquisition-site effects while preserving diagnostic information. Graph Neural Network (GNN) models achieve comparable, though not consistently superior, performance compared with a Random Forest (RF) classifier, indicating that increased model complexity alone does not overcome the domain shifts present in multi-site neuroimaging data. We introduced Spatial Graph Spectral Entropy (SGSE) to characterise the spatial organisation of graph-frequency variation in functional connectivity. The resulting entropy representation reveals a reproducible two-cluster organisation, termed Dynamic and Stable, which persists across different dimensionality-reduction settings. The two clusters exhibit distinct internal dimensionality and different patterns of diagnostic information. Dynamic is characterised by widespread reductions in local spectral entropy in ASD, whereas Stable exhibits more distributed alterations in functional connectivity. Subsequent classification analyses demonstrate that these clusters differ primarily in how diagnostic information is organised rather than in its overall amount. Finally, an error-focused analysis investigates subjects that remain difficult to classify across successive classification stages. The results indicate that these cases cannot be explained by a single set of strong connectivity abnormalities, but are consistent with greater inter-individual heterogeneity.
en
Additional information:
Arbeit an der Bibliothek noch nicht eingelangt - Daten nicht geprüft Abweichender Titel nach Übersetzung der Verfasserin/des Verfassers