4,068 posters, 219 videos, 101 topics, 69 sessions, 14,365 authors, 6,128 institutions

Alzheimer's Association International Conference 2026

July 12-15, 2026 | London, United Kingdom, and Online

1102

Hierarchical MRI-EHR Framework For Alzheimer’s And Mixed AD/ Vascular Dementia Subtyping

Background: Distinguishing Alzheimer’s disease (AD) from mixed or vascular cognitive

syndromes is challenging in routine memory-clinic practice. This distinction is increasingly

important because of knowledge of modifiable vascular risk factors and emerging disease

modifying AD therapies require accurate case selection, while vascular burden may influence

treatment safety and remains clinically actionable.

Methods: We used structural magnetic resonance imaging (MRI) volumetrics, white

matter hyperintensities (WMH) segmentation, radiology-report-derived features, demographic

variables, Mini-Mental State Examination (MMSE) scores and clinical comorbidities from

a cohort drawn from South London and Maudsley memory services to classify diagnoses

into AD, Mixed and Non-AD groups. We compared four XGBoost-based models: a direct

three-class model without WMH, a direct three-class model with WMH, a hierarchical model

without WMH and a hierarchical model with WMH. The main analysis used one available

baseline MRI per participant, XGBoost internal missing-value handling, and participant

grouped nested cross-validation with 5 outer folds and 10 inner folds. Model interpretability

was assessed using gain-based feature importance and supplementary SHAP-style dependence

plots.

Results: The baseline participant-level cohort included 3,963 participants. The direct

three-class model with WMH had the highest point estimates for macro-AUC (0.760, 95%

CI 0.749–0.773), macro-F1 (0.569, 95% CI 0.553–0.584), and balanced accuracy (0.582, 95%

CI 0.567–0.599). The hierarchical model with WMH had the highest point estimate for

overall accuracy (0.598, 95% CI 0.583–0.613). Adding WMH produced consistent but modest

improvements across both modelling strategies.

Conclusions: In this real-world memory-clinic cohort, the direct three-class model

with WMH had the strongest point estimates for class-balanced performance, whereas the

hierarchical model with WMH had the highest point estimate for overall accuracy and retained

a clinically interpretable two-stage structure. Explainability analyses highlighted clinically

plausible vascular, neurodegenerative and cognitive signals, with WMH and ischaemic report

features particularly relevant to Mixed classification. WMH added modest but consistent

1information, supporting further validation of vascular-burden-aware models as decision

support tools in routine dementia assessment.

Keywords: Alzheimer’s disease; mixed dementia; vascular dementia; memory clinic;