Modeling Temporal Dependencies and Feature Interactions Reveal Novel Clinical and Molecular Insights into Alzheimer's Disease Progression

Abstract

ObjectiveAlzheimers Disease remains a major public health challenge, requiring insights into feature interactions and temporal trends of feature importance. Community-wide data science competitions such as the TADPOLE Challenge provide platforms to benchmark predictive models using ADNI datasets. While top-performing models achieve accurate predictions, they often leave mechanistic questions unresolved. We introduce a framework that separately models static feature interactions and temporal dynamics, enabling complementary insights from longitudinal AD data. Materials and MethodsWe analyzed XGBoost with TreeSHAP for feature interactions and RNN-AD with Integrated Gradients for temporal trends. This two-branch design allows XGBoost to capture nonlinear cross-sectional interactions, while RNN captures evolving, time-dependent influences. The attributions are fused into a combined importance map. ResultsOur framework showed agreement between TreeSHAP and IG, highlighting FAQ, CDRSB, ADAS13, MMSE, and RAVLT variants as the most consistently important features across both branches. Temporal attribution analysis revealed stage-dependent trends: in CN, features such as DX:CN, FAQ, and RAVLT_immediate increased in importance with longer prediction horizons; in MCI, MidTemp and WholeBrain gained importance; and in AD, FAQ remained dominant. Feature-interaction analysis identified strong clinical-clinical interactions and secondary clinical-molecular interactions involving hippocampal and entorhinal volumes. DiscussionCombining interaction and temporal trends showed that RAVLT_immediate, FAQ, and DX-based features were the only markers consistently influential across both dimensions, indicating stable, cross-validated predictors of Alzheimers disease progression. Feature importance in AD prediction is dynamic, with early-time features often most influential. These insights support personalized monitoring, adaptive modeling, and mechanistic interpretability, enhancing patient-specific interventions and trial design. ConclusionThis work highlights feature interactions and temporal trends in AD prediction models, offering insights for personalized treatments and patient-specific trial designs. Our framework provides stable, cross-validated explanations that unify structural and temporal importance, enhancing trustworthiness and mechanistic interpretability in AD modeling.