Structural connectome and machine learning for Alzheimer’s disease detection

Abstract

AbstractBackgroundAlzheimer's disease (AD) is hypothesised as a disconnection syndrome where degenerating white matter fibre bundles leads to deterioration in the integration and communication between brain regions. Connectomics allows the study of in vivo brain connectivity and elucidates how the disease changes the brain network. Although some studies have shown evidence of alteration of structural connectivity between AD and cognitively normal individuals (CN), a large proportion of research focused on functional connectomics in AD. Emerging connectomics studies explored the use of machine learning (ML) to distinguish brain structural connectivity differences in AD with promising results. This project aims to identify changes in the structural connectome using novel image processing techniques to generate network metrics and utilise ML to classify between AD and CN.MethodWe examined data from 143 age‐matched subjects (AD mean: 71.1 ± 2.79 and CN mean: 71.09 ± 2.72) from the Alzheimer's Disease Neuroimaging Initiative cohort 2 (ADNI2). We used magnetic resonance images (T1‐weighted and diffusion‐weighted images) combined with the latest state‐of‐the‐art imaging processing tools to generate structural connectomes. Relevant network metrics were used to measure and compare brain connectivity, while ML algorithms were used to distinguish network metrics between AD and CN.ResultWe found significant connectivity changes in clustering coefficient (p < 0.05), normalised degree variance (p < 0.0001), hierarchical complexity (p < 0.005) and rich club (p < 0.0001) in AD (table 1). We also established and compared classification performances within our ML model. Random forest yielded sensitivity of 53.06% and specificity of 82.98% (table 2) for imbalanced data (AD=49, CN=94). On balanced data (AD=CN=49), the model was 81.63% specific and 69.39% sensitive (table 3) in detecting AD.ConclusionThe results show the feasibility of a connectome analysis of structural imaging combining the latest network metrics with ML for AD detection. While previous ML studies achieved promising results with balanced data, we reported both balanced and imbalanced models. As real‐world AD data are more likely to be imbalanced, the lower performance of the ML models on imbalanced data suggests that further improvement is needed for clinical implementation.