Global segregation and integration are associated with disease progression
Figure 2 shows the histograms of each subject group's mean cluster coefficient and global efficiency. Both indices for ADNI displayed an upward trend from the CN- to the Alzheimer’s disease stage. They were statistically different between CN- and EMCI, CN + and EMCI, EMCI and LMCI, and LMCI and AD (P < 0.05). In HABS, there is a significant difference between CN- and MCI, as well as CN + and MCI (P < 0.01). However, when comparing the CN- and CN+, neither ADNI nor HABS shows a difference in the clustering coefficients or global efficiency. This suggests that the global topological features are sensitive to the staging but not the change in a prodromal stage typically associated with amyloid-beta deposition.
Individual connectivity strengths are correlated with neuropsychological tests
Pearson’s correlation coefficients between the subject's mean Z-scores, global mean SUVR, and corresponding MMSE scores are shown in Fig. 3. For ADNI and HABS, Z-scores were inversely proportional to MMSE with values of − 0.5977 and − 0.3873, respectively. SUVR was likewise found to be inversely proportional to MMSE, with similar correlations. All correlations between neuropsychological scores were statistically significant (P < 0.01).
Topological features are better at staging than the composite SUVR
The tau accumulation pattern of each subject group was correlated to the Braak stages. Composite SUVR at various stages increases with the disease progression, albeit its capability in early MCI differentiation was restricted. The absolute mean Z-score of seven functional clusters for ADNI and HABS was shown in Fig. 4 and Fig. 5. Connectivity strength measures separated different disease groups better than SUVR. Each functional cluster connection can differentiate between the four subject groups (see Fig. 4). The biomarkers distinguished CN + from EMCI, EMCI from LMCI, and LMCI from AD, were intra-connectivity strength within MTL, ECN, and LAN. MTL, ECN, VIS. The intra-connectivity strength within LAN was most effective in distinguishing the CN + and EMCI groups. Furthermore, the Z-scores and SUVR of the two CN subgroups (CN- and CN+, Fig. 5) were compared. Although Z-scores perform slightly worse than other staging tasks, MTL, DMN, and VIS outperform SUVR (P < 0.01). For the validation cohort, again MTL, ECN, VIS and DMN contain the most significant biomarkers for distinguishing CN + and MCI stages. MTL, CCN, and VIS perform better than SUVR in differentiating CN- and CN+ (P < 0.01), whereas the other clusters do only marginally better. Cohen's effect sizes were further calculated for both ADNI and HABS cohorts. Z-scores result in larger effect sizes observed in all comparisons, indicating that using Z-scores to differentiate disease stages is more effective than SUVR (Supplementary Table S2, S3). In summary, SUVR was worse at identifying disease stages and recognizing early indications.
MTL and VIS as hubs identify early Alzheimer’s disease signs
Figure 6 depicts the unidirectional connections between the seven functional clusters, each represented by a 7 × 7 truncated matrix. The unidirectional connections between functional clusters in ADNI become increasingly complex as the stage progresses from CN to AD. There is a difference in connection abnormalities at CN- and CN + stages, although not statistically significant. The connections associated with MTL demonstrated a higher degree of connectivity in the EMCI stage (from MTL to VIS with Z-scores of 2.16, 5.49, and 9.96 in the EMCI, LMCI, and AD stages, respectively). For HABS, the connectivity also tends to be complex from CN- to MCI stage. The connectivity from VIS to ECN, DMN, and SM showed the most significant anomalies from CN- to CN+, with Z-scores of 2.41, 2.78, and 2.89, respectively. The connections from VIS, MTL, and CCN to other clusters are significantly abnormal from CN + to MCI, with most Z-scores greater than 2.16. This demonstrates that MTL and VIS abnormality hubs are indications of Alzheimer’s disease progression which is congruent with the clusters discovered in the previous section’s connectivity analysis.
Functional cluster connectivity pattern reveals patient heterogeneity
For ADNI, the mean Pearson’s correlation coefficients for subjects in the CN+, EMCI, LMCI, and AD groups were 0.9792 ± 0.00, 0.9504 ± 0.0085, 0.9408 ± 0.0247, 0.8663 ± 0.0572 and 0.7669 ± 0.0704, respectively. Similarly, for HABS, the mean Pearson’s correlation coefficients for CN + and MCI subjects were 0.9706 ± 0.00, 0.9183 ± 0.0324 and 0.8166 ± 0.0463, respectively. This indicates that heterogeneity among patients increases as the disease progress.
Pearson’s correlation coefficients for ADNI were 0.53, 0.49, 0.52, and 0.42 (P < 0.01), between the mean individual-level difference network difmean_individual and group-level difference network difmean_group for the four patient groups (CN+, EMCI, LMCI, and AD), respectively. However, the individual network for each patient has a low correlation with the group-level difference network difmean_group (CN+:0.1585 ± 0.0764, EMCI:0.1756 ± 0.0641, LMCI:0.1766 ± 0.1289, AD:0.2031 ± 0.0862). Similar findings were found for HABS, implying that each subject contributes to group-level differences differently.
Change of abnormal connectivity is associated with disease progression
Figure 7 shows the change of abnormal connectivity and SUVR in the cognitive control network for the two subjects whose disease status converted. Figure 7A shows the changes in abnormally connected edges before and after conversion from CN- to EMCI. After conversion, the global SUVR for these patients decreased (-2.9%) while the number of abnormal edges increased (absolute sum Z-score 48.9%). Figure 7B shows a patient who has converted from CN + to LMCI, with the change in connectivity (absolute sum Z-score 102.5%) being more noticeable than the change in SUVR (12.8%). This suggests that connectivity strength at the baseline scan may be more closely associated with forthcoming tau accumulation at the subject level.