Quantitatively modelling the phenomenology of Alzheimer's-type dementia using frame-theoretic semantics and psycholinguistic norms

Background The phenomenology of Alzheimer's disease—the “what is it like?” of the condition—has previously been approached via either quantitative analyses of granular linguistic features or qualitative analyses of lived experience. Objective What is lacking is an empirically credible framework that captures high-level qualitative insight by way of cognitive models rather than surface linguistic features. Methods Frame semantics and a 1.19 million-token psycholinguistic word norm lookup were applied to two topic-matched corpora: the DementiaBank Pitt Cookie Theft recordings and a COVID-era diary corpus combining CoSoWELL controls with Dementia Diaries entries. Gradient-boosted classifiers under participant-level cross-validation were compared against Fraser et al.'s 52-feature benchmark and statistically evaluated. Results A combined classifier contributed information statistically independent of the Fraser set on both corpora. On Pitt, four variance-based features defined a variance-collapse signature, with imageability being most significant; the within-document standard deviation of imageability replicated on COVID. Twelve semantic-frame comparisons survived Holm correction; Kinship was under-used on both corpora, giving a second cross-corpus anchor alongside imageability. A four-feature minimal classifier drawing equally from both feature families achieved AUC 0.748 on Pitt and 0.781 on COVID. Conclusions Alzheimer's-type dementia is characterized not by a shift in the mean of any linguistic dimension but by a flattening of the phenomenological range across which a speaker moves when describing and experiencing the world. Because the diagnostic labels available in both corpora are coarse, the extent to which this flattening reflects a unitary phenomenology of Alzheimer's disease rather than heterogeneity within the diagnostic category remains to be established.