Schedule of Events | Search Abstracts | Invited Symposia | Symposia | Poster Sessions | Data Blitz

Explainable Multimodal Models for Cognitive Aging: SHAP Analyses Reveal Clinically Relevant Nonlinear Structural and Graph-Theoretical Network Features

Poster Session E - Monday, March 9, 2026, 2:30 – 4:30 pm PDT, Fairview/Kitsilano Ballroom
Also presenting in Data Blitz Session 4 - Saturday, March 7, 2026, 10:30 am – 12:00 pm PST, Salon F.

Héctor Manuel Cárdenas Castro¹ (), Jose Antonio Cantoral Ceballos¹, Ricardo Caraza^1,3, Luis Angel Trejo Rodriguez¹, Alejandro Castañeda Miranda²; ¹Instituto Tecnologico y de Estudios Superiores de Monterrey, ²Universidad Aeronáutica de Querétaro, ³Centro de Neurociencias Cognitivas

Understanding cognitive aging (a continuum from preserved cognition to neurodegeneration) demands biomarkers that link regional structure, network organization and clinical staging. We introduce an explainable multimodal deep-learning framework to resolve black-box uncertainty and map continuous and categorical aging phenotypes. Methods: T1-weighted volumetrics and resting-state fMRI graph-theory metrics were combined from three public cohorts: OASIS3, NIMH-ds002415, and ANT-ds001907 (N = 959). Extracted features included cortical and subcortical volumes, and graph metrics (characteristic path length, global/local efficiency, modularity) computed within network atlas. Models were stacked neural networks producing a brain-age regressor and a dementia classifier; training used nested cross-validation and class-balanced sampling to reduce overfitting and imbalance. SHAP was applied to quantify main effects, nonlinear feature interactions and subgroup-specific contributions. Results: The dementia classifier achieved strong performance (weighted F1 ≈ 0.93 with precision and recall) and the brain-age model yielded low error in controls (MAE ≈ 0.7 years). Brain-Age Gap (BAG) correlated moderately with clinical severity (CDR; r ≈ 0.32), supporting BAG as a continuous staging marker. SHAP indicated structural dominance: ventricular enlargement and focal temporal/insular volumes were primary drivers of classification, while connectivity metrics acted as modulatory signals that refined staging. SHAP distributions and subgroup analyses revealed heterogeneous, bell-shaped and context-dependent contributions, implying many imaging–clinical relationships are nonlinear and interaction-driven instead of strictly additive. Conclusion: Results argue for a shift: beyond linear biomarker mapping. XAI methods (SHAP) coupled with Complex Systems Theory may capture emergent dynamics, compensatory reserve and critical transitions, offering interpretable, clinically actionable monitoring of cognitive aging.

Topic Area: METHODS: Neuroimaging