Cross‐National Study to Operationalize Amyloid, Tau, Neurodegeneration, and Vascular Contributions to Dementia

TL;DR

This study compares methods to diagnose dementia using brain scans and biomarkers in U.S. and Korean populations, finding that continuous measures work best for research.

Contribution

The paper introduces a cross-national comparison of multimodal biomarker operationalization for dementia diagnosis, highlighting regional differences in model performance.

Findings

Models with continuous ATNV measures explained more dementia diagnosis variability in the U.S. cohort.

Korean models showed higher predictive power when including binary N and binary A.

Continuous variables are optimal for research, while binary variables may enhance clinical predictions in specific regions.

Abstract

Recent diagnosis and staging criteria have been proposed for AD based on amyloid(“A”), tau(“T”), and neurodegeneration(“N”), with consideration of comorbid vascular(“V”) pathology. However, challenges remain in their application and interpretation for research and clinical use. Additionally, current criteria have been largely informed by data derived from samples of highly educated, non‐Hispanic White cohorts in the US. We compared methods to operationalize multimodal ATNV measures across two cohorts to meaningfully inform future research and clinical practice.

Participants with amyloid PET, tau PET, and brain MRI were included from two prospective cohort studies with similar study designs: ADNI3 in the US and KBASE in Korea. ATNV measures were operationalized as continuous (A=centiloid; T=meta‐temporal ROI; N = AD signature region cortical thickness; V=white matter hyperintensity volume) and as binary (+/‐) variables by replicating approaches applied in other cohorts. Clinical diagnoses of cognitively unimpaired (CU), mild cognitive impairment (MCI), and dementia were determined by experts at multidisciplinary consensus conferences. Parallel cross‐sectional analyses were performed within each cohort. Multivariate logistic regressions with ATNV modeled as continuous or binary predictors (16 possible combinations) adjusting for age, sex, and education were performed to compare model fit and predictive power for distinguishing participants with dementia versus CU.

A total of 508 participants in ADNI (mean age=71±7, female=55%, education(yrs)=16.5±2.3) and 165 in KBASE (n = 165; age=73±8, female=64%, education(yrs)=11.0±4.6) were included (Table 1). In ADNI, the continuous ATNV and continuous ATN+binary V explained the most variance in dementia diagnosis (highest pseudo‐R2, lowest AIC/BIC; Table 2). In KBASE, models had higher overall predictive power for distinguishing dementia. The best‐performing models included continuous A and T, binary N, and continuous V, or binary A, continuous T, binary N, and continuous V (Table 3).

Models including ATNV were highly predictive of dementia in both cohorts. Models with predominantly continuous variables explained more variability in dementia diagnosis, and are optimal for research use due to their simplicity. In Korea, binary N and binary A increased predictive power. Ongoing analyses will assess the relative contribution of ATNV to cognitive performance and diagnosis in each cohort to better inform appropriate research and clinical use in cross‐national cohorts.

Figures