Integrating individualized connectome with amyloid pathology improves predictive modeling of future cognitive decline

Abstract

Background: The deposition of amyloid-β (Aβ) in the human brain is a hallmark of Alzheimer's disease and is associated with cognitive decline. Aβ pathology is traditionally assessed at the whole-brain level across neocortical regions using positron emission tomography (PET). However, these measures often show weak associations with future cognitive impairment. A more sensitive pathology metric is needed to quantify early Aβ burden and better predict cognitive decline. Here, we aim to develop a network-based metric of Aβ burden to improve early prediction of cognitive decline in aging populations.

Methods: We integrated subject-specific brain connectome information with Aβ-PET measures to construct a network-based metric of Aβ burden. Cross-validated predictive modeling was used to evaluate the performance of this metric in predicting longitudinal cognitive decline. Furthermore, we identified a neuropathological signature pattern linked to future cognitive decline, and we validated this pattern in an independent cohort.

Results: Our results demonstrate that incorporating individualized structural connectome, but not functional connectome, information into Aβ measures enhances predictive performance for prospective cognitive decline. The identified neuropathological signature pattern is reproducible across cohorts.

Conclusion: These findings advance our understanding of the spatial patterns of Aβ pathology and its relationship to brain networks, highlighting the potential of connectome-informed network-based metrics for Aβ-PET imaging in identifying individuals at higher risk of cognitive decline.

Fig. 1. Regional amyloid-β standardized uptake value ratio (SUVR) and network-based amyloid-β pathology (NAP).

In addition to regional SUVR, the proposed NAP scores incorporate connectome information by using either a connectivity-weighted or centrality-scaled approach. Here, we show a six-region toy network. a Illustration of regional SUVR deposition A_i in region i (i = 1,…, 6). b Connectivity-weighted NAP score quantified the influence of amyloid-β (Aβ) deposition within the connected networks. c Centrality-scaled score quantifies the Aβ deposition in the region and scaled it by its centrality in the connectome. Network connectome information can be generated based on either structural connectome (SC) (d) or the functional connectome (FC) (g). d SC is derived from tractography on diffusion tensor imaging data, where the SC matrix was denoted as $C_{ij}^S$. e SC-based connectivity-weighted NAP score of region j quantifies both the Aβ in region j and a weighted sum of Aβ in all other regions based on connectivity values, where the whole process was denoted as matrix multiplication of $A_i\times \left(C_{ij}^S+I\right)$. f SC-based centrality-scaled NAP score of region i is quantified by a Hadamard product of regional SUVR measure and connectivity centrality $A_i\odot \left({\sum }_j^6{C}_{ij}^s\right)$. g FC is computed from statistical dependency on functional MRI data. For the connectivity-weighted (h) and centrality-scaled pathology (i), FC-based NAP scores are computed in the same way as the SC-based approach, but with FC denoted as $C_{ij}^F$.

Fig. 2. Cross-validated predictive modeling analysis overview.

The figure illustrates an example of using connectivity-weighted network-based amyloid-β pathology (NAP) with a functional-based connectome. a Feature selection on the input of pathology models and connectome. Regional standardized uptake value ratio (SUVR) or NAP (either connectivity-weighted or centrality-scaled) can be used as input. Regional SUVR measures do not need the input of connectome information, and the centrality-only model has only connectome information (serving as a benchmark for NAP measures). A feature is selected if it shows a negative correlation with the cognitive change residual score at an uncorrected p < 0.05 in the training set. b Computation of longitudinal cognitive change scores. Subjects are randomly split into a training set and a test set. In the training set, we regress out covariate confounds from the longitudinal cognitive change. The covariate-related variability is also removed from the test set using the same weights. c Regression of selected amyloid pathology features in the train set and prediction of cognitive decline in the test set. The sum of selected features is regressed against cognitive change residual scores from the training sample to obtain model weights. The obtained weights are then applied to the features from the test set to predict cognitive decline scores, which were then compared to actual scores to evaluate model prediction performance.

Fig. 3. Performance of cognitive decline prediction using different pathology models across cross-validation repetitions.

a Distributions of differences in correlation coefficients between predicted and actual cognitive decline scores across different pathology models. Prediction performance was compared across cross-validation repetitions with the regional standardized uptake value ratio (SUVR) model as the reference. Incorporating structural connectome (SC) with connectivity-weighted network-based amyloid-β pathology (NAP) achieved the highest performance, and outperformed the prediction performance of regional SUVR measures. b Comparison of the ratio of root mean squared error (RMSE) performance between pathology models. Compared to regional SUVR measures, connectivity-weighted NAP measure of SC had smaller RMSE, whereas connectivity-weighted NAP measures of functional connectome (FC) had higher RMSE. The analysis included seventy-seven participants for whom all data were available.

Fig. 4. Regional standardized uptake value ratios (SUVR) and network-based amyloid-β pathology (NAP) maps.

a Group averaged structural connectome (SC) connectivity-weighted NAP scores. b Correlation coefficient between cognitive decline and SC connectivity-weighted NAP scores across cortical regions of interest (ROIs) (Spearman partial correlation, controlling for age, sex, education, baseline cognition, and hemispherical mean cortical thickness). c Correlation coefficient between cognitive decline and regional SUVR scores across cortical ROIs. NAP scores have higher values in basal cortex regions (a). Individual variability in NAP scores in basal cortex regions and the left dorsolateral prefrontal cortex scales with cognitive decline (b), showing a stronger negative correlation with subsequent cognitive change compared to regional SUVR scores (c). The analysis included seventy-seven participants for whom all data were available.

Fig. 5. The amyloid-β (Aβ) neuropathological signatures of future cognitive decline and external validation.

a Regions of interest (ROIs) whose regional Aβ standardized uptake value ratio (SUVR) measures were identified as features of future cognitive decline. This pattern of pathology signature was derived from the internal CogRes/RANN dataset. In external validation, features within the same ROIs pattern of ADNI participants were used to predict cognitive decline in dMemory, and subsequently compared to the actual decline score. The predicted cognitive change significantly correlated with the actual score (Pearson correlation; two-sided), and the relationship is significantly higher than the results obtained by random ROI selection (Randomization test; one-sided), demonstrating regional specificity. b Results of Aβ neuropathological signatures using connectivity-weighted network-based Aβ pathology (NAP) measures. c Results of Aβ neuropathological signatures using centrality-scaled NAP measures. The color indicates the network label of each ROI based on the Schaefer atlas. The analysis included seventy-two participants for whom all data were available. Cont: cognitive control network, DorsAttn: dorsal attention network, VisCent: visual central network, VisPeri: visual peripheral network, SalVentAttn: salience/ventral attention network, Default: default mode network.

Fig. 6. Correlation results in the relationship between pathology features in each region of interest (ROI) and longitudinal cognitive decline.

a For participants in the CogRes/RANN study, we performed a Spearman partial correlation analysis between pathology features in each ROI and longitudinal cognitive changes, controlling for age, sex, education, baseline cognition, and mean cortical thickness of each hemisphere. Pathology features were characterized by either regional standardized uptake value ratio (SUVR) or network-based amyloid-β pathology (NAP). b Same approach was performed for participants in the ADNI study. Compared to regional SUVR measures, NAP scores demonstrated a significantly stronger negative relationship with cognition change in both studies (In the CogRes/RANN study: connectivity-weighted NAP: p < 6.06e-21; centrality-scaled NAP: p < 1.19e-20; In the ADNI study: connectivity-weighted NAP: p < 1.01e-05; centrality-scaled NAP: p < 1.30e-03). Cognition was assessed using global cognition in the CogRes/RANN study and dMemory in the ADNI study. The analysis included seventy-seven participants in the CogRes/RANN study and seventy-two participants in the ADNI study. (**p < 0.01; two-sample Student’s t-test).