Individual Brain Identification via Functional Connectivity: From Classical Methods to Graph Attention Networks

Abstract

Understanding the human brain as an individualised biological system has emerged as a central goal of contemporary neuroscience. This research investigates brain fingerprinting, the phenomenon whereby an individual’s functional connectome serves as a stable and unique biological signature. Despite high identification accuracies achieved using linear metrics such as Pearson correlation, current methods often fail to capture complex non-linear dependencies and discard vital non-Euclidean graph topology by flattening the connectome into vectors. This thesis proposes a systematic, three-stage experimental framework to address these limitations, evaluating models of increasing expressive capacity: a Pearson correlation base line, classical machine learning classifiers formulated as pairwise verification tasks, and deep Siamese architectures utilising Graph Neural Networks (GNNs) governed by an InfoNCE con trastive loss. Experiments are conducted on resting-state fMRI data from the HCP-Young Adult and HCP-Aging cohorts, with atlas resolution (Schaefer-100 vs. Schaefer-400) varied systematically across both datasets. Results demonstrate that supervised discriminative approaches significantly outperform unsupervised baselines. The Siamese Graph Attention Network (GAT) is the strongest per former across all conditions, achieving 100% identification accuracy on HCP-Aging at fine grained resolution. Its learned attention maps provide neurobiological interpretability by naturally prioritising the frontoparietal (Control) network, independently corroborating es tablished theories on inter-individual variability. Model rankings remain consistent across cohorts, with performance differences attributable to demographic and acquisition factors. Finer parcellations consistently benefit graph-based models while penalising linear classifiers, providing empirical guidance for future methodological choices. By establishing a robust, topologically-aware, and interpretable framework for brain fingerprinting, this work contributes to the advancement of individualised clinical tools and precision medicine.