Graph-theoretical mapping of cortical and subcortical network alterations in preclinical neurodegeneration

Abstract

Structural disruptions in brain connectivity often precede the overt symptoms of neurodegenerative and neurodevelopmental disorders. Graph-theoretical metrics derived from structural magnetic resonance imaging (MRI) and diffusion tensor imaging (DTI) offer a non-invasive window into these changes. When combined with machine learning (ML), these features may improve early diagnostic precision. This review aimed to evaluate the utility of structural network metrics and supervised ML algorithms in identifying preclinical brain network alterations in neurological conditions. A comprehensive search of the PubMed, Scopus, and Web of Science databases was conducted (updated April 2025). The inclusion criteria targeted studies using graph-theoretical analysis of T1-weighted and DTI-derived networks in preclinical studies incorporating interpretable supervised ML algorithms, specifically the Support Vector Machine (SVM) or Random Forest (RF). While this focus ensured methodological consistency, it may have excluded other high-performing models (deep learning and ensemble algorithms). Data extraction included imaging protocols, network features, ML performance and cognitive correlations. Fifty-seven studies met the inclusion criteria of this review. Global efficiency, clustering coefficient, and path length are consistently altered in preclinical Alzheimer's disease (AD), Parkinson's disease (PD), and small vessel disease (SVD). The SVM and RF models achieved classification accuracies of up to 93%, with AUCs ranging from 0.78 to 0.95. Strong correlations were observed between network metrics and cognitive scores (global efficiency, r = 0.65, P < 0.001). Harmonization methods were inconsistently applied across the datasets, affecting their reproducibility. Structural connectomics combined with ML enables the identification of early brain network disruptions. Reduced efficiency and increased path length are key markers of preclinical conditions. These findings support the translational potential of ML-integrated network analysis for early risk stratification and targeted intervention.