Fig. 1

Schematic illustration of the structural covariance network and prediction model. A PLS regression was first conducted between the selected seed region and the remaining voxels throughout the whole brain, and the top 3% of voxels with the largest correlation (weight) with the seed region were retained to build a grey matter structural covariance network. Then, the structural covariance network score of each individual was calculated. B The prediction models of RF were established based on the baseline score maps of the covariance network. Abbreviations: SCN, structural covariance network; PLS, partial least squares; RF, random forest; sNC, stable normal cognition; pNC, progressive normal cognition; ROC, receiver operating characteristic curves