Dynamic network connectivity predicts subjective cognitive decline: the Sino-Longitudinal Cognitive impairment and dementia study

Abstract

Subjective cognitive decline (SCD) is the preclinical stage of Alzheimer's disease (AD), the most common neurodegenerative disease in the elderly. We collected resting-state functional MRI data and applied novel graph-theoretical analyses to investigate the dynamic spatiotemporal cerebral connectivities in 63 individuals with SCD and 67 normal controls (NC). Temporal flexibility and spatiotemporal diversity were mapped to reflect dynamic time-varying functional interactions among the brain regions within and outside communities. Temporal flexibility indicates how frequently a brain region interacts with regions of other communities across time; spatiotemporal diversity describes how evenly a brain region interacts with regions belonging to other communities. SCD and NC differed in large-scale brain dynamics characterized by the two measures, which, with support vector machine, demonstrated higher classification accuracies than conventional static parameters and structural metrics. The findings characterize dynamic network dysfunction that may serve as a biomarker of the preclinical stage of AD.

Keywords: Alzheimer’s disease; Resting-state functional MRI; Spatiotemporal diversity; Subjective cognitive decline; Temporal flexibility.

Fig. 1

The computation of dynamic graph theoretic parameters. A sliding window approach was used to quantify the time-varying changes in the community structure of intrinsic functional connectivity. The temporal co-occurrence matrix was based on the Louvain community detection algorithm. The temporal flexibility and spatiotemporal diversity were computed by an optimized community detection algorithm

Fig. 2

Algorithms for group classification. We used parameters of dynamic graph theory, static graph theory and gray matter volume as the features of SVM classifier, respectively. The Lasso sparse formula was used for feature selection. Ten-fold cross validation was used to obtain the accuracy of the classifier. We adjusted the penalty term parameters of the Lasso sparse formula based on the accuracy to obtain the optimal classification accuracy for each feature

Fig. 3

The receiver operating characteristic (ROC) curves of neural features. The two dynamic connectivity measures showed the highest prediction accuracy

Fig. 4

Top 10 regions with the largest weight of dynamic graph theoretic features. a Regions in red/blue each showed higher/lower mean temporal flexibility in SCD than in NC. b Bar plots of mean value of temporal flexibility of the top 10 brain regions, *** for P < 0.001; two sample t test. c Regions in red/blue each showed higher/lower mean spatiotemporal diversity in SCD than in NC. d Bar plots of mean value of spatiotemporal diversity of the top 10 brain regions, * for P < 0.05, ** for P < 0.01, *** for P < 0.001; two sample t test

Fig. 5

Behavioral correlation with dynamic graph theoretic parameters (FDR corrected, P < 0.05) across NC and SCD. NC: normal controls; SCD: subjective cognitive decline. AFT: animal fluency test; BNT: Boston naming test; AVLT: auditory verbal learning test