Predicting Cognitive Level with Resting-state Functional Connectivity for Alzheimer's Disease

Neuroimaging is an effective and objective method to assess cognitive levels, minimizing bias from subjective measurements and identifying relevant neural markers. In this study, we proposed a novel data-driven cognitive prediction framework called ensemble connectome-based relevance-vectors-regression predictive modeling (ECRPM). After obtaining the relevant resting-state functional magnetic resonance imaging (fMRI), we applied the previously proposed joint human connectome project multi-modal parcellation (J-HCPMMP) method to derive the resting-state functional connectivity (rs-FC), which is an undirected graph represents 360 brain regions, then extracted connections between brain regions related to cognition using the significant edge selection (SES) method. Following dimensionality reduction, the feature data was combined with the connectome and cognitive measurements using the relevance vector regression (RVR) method. To enhance the robustness and predictive power of the model, we used an ensemble learning framework, simultaneously trained multiple homogeneous models and assigned appropriate weights to each base model based on their predictive abilities. Using this framework to predict ADASt3 scores, we found strong correlations between the predicted and observed cognitive scores. The positive network was more effective in predicting ADASt3 scores (r=O.780). To further verify the model's generalization ability, we employed k-fold cross-validation and leave-one-out cross-validation methods. The results demonstrated the framework's strong cognitive prediction ability. Finally, analysis of the feature data revealed core brain regions related to cognitive decline, particularly the entorhinal cortex (EC) region, which aligns closely with previous studies.