AT(N)-Anchored Multimodal Deep Learning for Alzheimer’s Staging: A Taxonomy and Continuous Burden Index for Low-Resource Settings

The global burden of dementia is heavily concentrated in low- and middle-income countries (LMICs), yet biological diagnosis of Alzheimer’s Disease (AD) relies on positron emission tomography (PET) and cerebrospinal fluid (CSF) analysis that are operationally inaccessible in these regions. The 2024 Alzheimer’s Association revised criteria recognize sufficiently accurate plasma assays,particularly p-tau217,as Core 1 biomarkers that may support biological diagnosis, offering a scientifically credible pathway to infrastructure-light biological detection and staging. However, the machine learning literature has expanded without principled mapping to the AT(N) biological framework or rigorous consideration of deployability constraints in resource limited settings. We present a systematic scoping review synthesizing 146 included full-text publications (2018–2026) across four multidisciplinary evidence streams, establishing a novel architectural taxonomy that maps deep learning architectural families directly to AT(N) biological targets and connects model inductive biases to specific biomarker categories. We audit major AD datasets, identify critical biases in demographic and biomarker representation, and map pervasive data pathologies—label scarcity, missing modalities, scanner heterogeneity, and population bias—to corresponding ML/DL remediation strategies. Finally, we argue for a shift from binary classification toward continuous, biologically anchored disease-burden scoring, proposing the Alzheimer’s Disease Burden Index (ADBI). The ADBI equations presented herein represent a theoretical and architectural framework designed to guide future prospective validation on multi-ethnic cohorts. This work provides a methodological roadmap for developing globally deployable, equity aware AD diagnostic models.