Abstract
Alzheimer's disease (AD) manifests with varying progression rates across individuals, necessitating the understanding of their intricate patterns of cognition decline that could contribute to effective strategies for risk monitoring. In this study, we propose an innovative interpretable population graph network framework for identifying rapid progressors of AD by utilizing patient information from electronic health-related records in the UK Biobank. To achieve this, we first created a patient similarity graph, in which each AD patient is represented as a node; and an edge is established by patient clinical characteristics distance. We used graph neural networks (GNNs) to predict rapid progressors of AD and created a GNN Explainer with SHAP analysis for interpretability. The proposed model demonstrates superior predictive performance over the existing benchmark approaches. We also revealed several clinical features significantly associated with the prediction, which can be used to aid in effective interventions for the progression of AD patients.
Introduction
Alzheimer’s Disease is notoriously heterogeneous; while some patients experience a slow decline over 15 years, others progress from mild cognitive impairment (MCI) to full-blown dementia or death within a much shorter window. Identifying these "Rapid Progressors" (RPs) is critical for clinical trial recruitment and personalized care, yet there is no universal consensus on how to define or predict them.
Current methods often rely on specific cognitive tests like the Mini-Mental State Examination (MMSE), which can be inconsistent or unavailable in standard medical records. This research addresses the challenge of using high-dimensional electronic health record (EHR) data—including comorbidities, lifestyle, genetics, and treatments—to build a predictive model that is both accurate and interpretable for clinical use.
Methodology
The core innovation of this study is the construction of a Population Graph. Instead of treating each patient as an isolated data point, the framework represents each patient as a node. An edge (connection) is established between two patients if their clinical characteristics are sufficiently similar.
Key Components:
- Hybrid Distance Metric: Patient similarity is determined by a weighted sum of Average Jaccard Similarity (for categorical data) and Chebyshev Distance (for numerical data).
- The "Buffer" Strategy: To define "rapid" vs. "slow" progression without the bias of an arbitrary cutoff, the authors introduced a "buffer" ratio. They calculated the mean progression time and excluded patients in the middle-ground window, creating two distinct classes (RPs vs. non-RPs).
- GNN Architectures: Three types of GNNs were evaluated: Graph Sample and Aggregate (SAGE), Graph Convolutional Networks (GCN), and Graph Attention Networks (GAT).
- Interpretability: The model integrates GNN Explainer with SHAP (SHapley Additive exPlanations) analysis, allowing the system to highlight which specific clinical features most influenced a prediction.
Experimental Results
The GNN-based approach demonstrated clear superiority over traditional models like Logistic Regression, Random Forest, and standard Deep Neural Networks (MLPs, BiLSTMs).
- MCI-AD Cohort: The SAGE model achieved an Accuracy of 0.929 and an AUROC of 0.976. In contrast, a standard Random Forest only managed an AUROC of 0.532.
- AD-Death Cohort: The GCN model was the top performer, reaching an Accuracy of 0.930 and an AUROC of 0.983.
Clinical Insights
The SHAP analysis revealed that the model identifies non-obvious predictors of rapid decline. Beyond neurological markers, the model highlighted systemic conditions and lifestyle factors. For instance, cardiovascular health (indicated by chest pain), mood swings, and high salt intake were strongly associated with faster progression.
Discussion and Future Work
The success of the population graph suggests that a patient’s risk is best understood by looking at "clinical neighbors." If a patient is surrounded by others with similar comorbidities who progressed rapidly, the GNN captures that relational signal.
While the results are promising, the authors acknowledge limitations such as the relatively small size of specific MCI-to-AD transition cohorts in the UK Biobank and the inherent latency in EHR diagnosis dates. Future work will involve testing the framework on multi-site datasets and implementing Federated Learning to enhance generalizability while maintaining patient privacy.
Summary updated