Discovering Invariant Patterns of Cognitive Decline Via an Automated Analysis of the Cookie Thief Picture Description Task

Anna Favaro, Najim Dehak, Thomas Thebaud, Jesús Villalba, Esther S. Oh, Laureano Moro-Velázquez

Abstract: The Cookie Thief task has been extensively adopted to uncover patterns characterizing Alzheimer's Disease (AD). Yet, how findings reported on this task generalize across different corpora remains unclear. The main objective of this study was to assess the cross-corpora robustness of 145 cognitive, linguistic, and acoustic features, conducting a statistical and correlation analysis with clinical scores. Thus, four corpora were considered. The experimental results indicate that cognitive features reflecting attention focus and linguistic features representing information units and vocabulary richness displayed a robust behavior across corpora and reported a strong correlation with clinical scores. However, paralinguistic features such as pause length and pitch variability reported a lack of homogeneous behavior. This study provides insights into inter-corpora dissimilarities and highlights which features can be used when applying models trained in one domain to an unseen cohort or domain.