Uncovering Alzheimer's Disease Progression via SDE-based Spatio-Temporal Graph Deep Learning on Longitudinal Brain Networks

Abstract

Identifying objective neuroimaging biomarkers to forecast Alzheimer's disease (AD) progression is crucial for timely intervention. However, this task remains challenging due to the complex dysfunctions in the spatio-temporal characteristics of underlying brain networks, which are often overlooked by existing methods. To address these limitations, we develop an interpretable spatio-temporal graph neural network framework to predict future AD progression, leveraging dual Stochastic Differential Equations (SDEs) to model the irregularly-sampled longitudinal functional magnetic resonance imaging (fMRI) data. We validate our approach on two independent cohorts, including the Open Access Series of Imaging Studies (OASIS-3) and the Alzheimer's Disease Neuroimaging Initiative (ADNI). Our framework effectively learns sparse regional and connective importance probabilities, enabling the identification of key brain circuit abnormalities associated with disease progression. Notably, we detect the parahippocampal cortex, prefrontal cortex, and parietal lobule as salient regions, with significant disruptions in the ventral attention, dorsal attention, and default mode networks. These abnormalities correlate strongly with longitudinal AD-related clinical symptoms. Moreover, our interpretability strategy reveals both established and novel neural systems-level and sex-specific biomarkers, offering new insights into the neurobiological mechanisms underlying AD progression. Our findings highlight the potential of spatio-temporal graph-based learning for early, individualized prediction of AD progression, even in the context of irregularly-sampled longitudinal imaging data.

Figures (19)

• Figure 1: An overview of our proposed model for AD progression prediction and biomarker interpretation. (a) The datasets consist of longitudinal fMRI scans, with up to six time points spanning from baseline to 105 months of follow-up. (b) The longitudinal fMRI scans are preprocessed and reconstructed to continuous signals for each ROI by the SDE method. The dynamic graphs were built by using longitudinal reconstructed fMRI signals. First, the connections between ROIs were quantified by measuring the correlation between their reconstructed signals. Next, the reconstructed signals in the ROIs, along with their connections, were combined with the proposed learnable importance probabilities to generate a sparse graph. The dynamic graphs thus consist of longitudinal sparse graphs. (c) The dynamic graphs are sent to our SDE-guided spatio-temporal GNN model to learn and evolve the longitudinal representation. The learned representation is fed into an MLP classifier to predict the disease progression. The importance probabilities on nodes and edges of sparse graphs provide the interpretation for the salient ROIs and the prominent disease-specific connections.
• Figure 2: Classification comparison of stable vs. progressive subjects in OASIS-3 between the state-of-the-art machine learning models and ours by using different preprocessing strategies including SDE, BrainODE, RNN, and Mean. The average classification accuracy, the area under the receiver operating characteristic curve (ROC-AUC), specificity, and sensitivity are reported under the 5-fold cross validation. The best performance was achieved by using our proposed method.
• Figure 3: Interpretation analysis in OASIS-3 cohort. (a) Interpreting top 20 selected salient ROIs in the progressive group across six models, each trained on an increasing number of timesteps. The title indicates the mean follow-up months of subjects. The color bar ranges from 0.3 to 1.0. The bright-yellow color indicates a high score, while dark-red color indicates a low score. (b) The significant difference of the interpreted most discriminative connections for distinguishing stable and progressive subjects was evaluated by two-sample t-tests with FDR corrected p-value $<$ 0.05. Here, the top 30 most discriminative ROI connections are visualized for interpretation over 6 different timesteps. The intra-network connections are colored based on the module itself and inter-network connections are colored in grey. (c) Neural system-level interpretation of the most discriminative connections was computed by averaging the absolute t value of most discriminative ROI connections reported between neural systems over 6 different timesteps. The dark-red color indicates significant connections. The non-significant connections are marked as white. VIS = Visual Network, SMN = Somatomotor Network, DAN = Dorsal Attention Network, VAN = Ventral Attention Network, LIM = Limbic Network, CON = Control Network, DMN = Default Mode Network.
• Figure 4: (a) Classification comparison of stable vs. progressive contrast in ADNI validation cohort between the state-of-the-art machine learning models and ours by using different preprocessing strategies including SDE, BrainODE, RNN, and Mean. (b) Classification performance of independent ADNI and OASIS-3 test cohort using all available timepoints. The average classification accuracy (Acc), the area under the receiver operating characteristic curve (ROC-AUC), specificity (Spe), and sensitivity (Sen) are reported under the 5-fold cross-validation.
• Figure 5: Reproducibility of interpretation in validation ADNI cohort. (a) Interpreting top 20 selected salient ROIs in the progressive group across six models, each trained on an increasing number of timesteps. The title indicates the mean follow-up months of subjects. The color bar ranges from 0.3 to 1.0. The bright-yellow color indicates a high score, while dark-red color indicates a low score. (b) The significant difference of the interpreted most discriminative connections for distinguishing stable and progressive subjects was evaluated by two-sample t-tests with FDR corrected p-value $<$ 0.05. Here, the top 30 most discriminative ROI connections are visualized for interpretation over 6 different timesteps. The intra-network connections are colored based on the module itself and inter-network connections are colored in grey. (c) Neural system-level interpretation of the most discriminative connections was computed by averaging the absolute t value of most discriminative ROI connections reported between neural systems over 6 different timesteps. The dark-red color indicates a high score. The non-significant connections are marked as white. VIS = Visual Network, SMN = Somatomotor Network, DAN = Dorsal Attention Network, VAN = Ventral Attention Network, LIM = Limbic Network, CON = Control Network, DMN = Default Mode Network.