S-038 | Language-based Alzheimer’s identification: A comparison with neuropsychological and neuroimaging features

Cognition, Behavior, and Memory
Author: Ivan Caro | Email: ivan.caro.strokes@gmail.com

Ivan Caro^1°2°, Gonzalo Pérez^1°2°, Joaquín Valdés Bize^3°,  Joaquín Ponferrada^1°, Franco Ferrante^1°2°, Lara Gauder^4°, Luciana Ferrer^4°,  Agustín Ibañez^1°2°, Andrea Slachevsky^1°, Adolfo M. García^1°

^1° Cognitive Neuroscience Center (CNC), Universidad de San Andrés, Buenos Aires, Argentina
^2° National Scientific and Technical Research Council (CONICET), Buenos Aires, Argentina
^3° Department of Psychiatry, School of Medicine, Pontificia Universidad Católica de Chile, Santiago, Chile
^4° Instituto de Investigación en Ciencias de la Computación (ICC), CONICET-UBA, Argentina, Departamento de Computación, Faculty of Exact and Natural Sciences, University of Buenos Aires (UBA), Argentina.

Automated word property (WP) analysis represents a powerful digital innovation to reveal scalable markers of Alzheimer’s disease (AD). Quantification of lexical features during oral production proves useful to detect AD cases and to predict cognitive outcomes and anatomical-functional brain patterns. However, no study has compared the discriminatory capacity of WP markers with standard neuropsychological and neuroimaging measures, casting doubts on the actual clinical contributions of the approach. This proof-of-concept study aims to compare the relative robustness of WP markers with features extracted from neuropsychological and brain assessments. We recruited 33 AD patients and 33 healthy controls from a carefully characterized cohort. All participants underwent verbal fluency tests, neuropsychological evaluations (tapping on general cognitive skills, attention, set-shifting, and working memory) and MRI scans. Separate machine learning classifiers were trained with (i) WP features from the fluency tasks, (ii) score from the neuropsychological tests, and (iii) volumetric features from the imaging protocol. Cross-validation results showed that patient identification was similar between WP features (AUC=.828) and both neuropsychological (AUC=.814) and imaging (AUC=.892) features (p-values>0.05). Overall, WP analyses seem non-inferior to standard diagnostic measures, reinforcing their value as a scalable, low-cost tool for dementia assessments.