Linguistic features from constrained spontaneous speech as potential markers of preclinical Alzheimer’s disease

Background Language impairment is an early sign of Alzheimer’s disease (AD) that may emerge before more obvious cognitive decline. Linguistic features derived from verbal responses to cognitive tests have shown promise for identifying mild cognitive impairment and predicting AD risk; however, their relationships with established AD biomarkers remain insufficiently characterized. Objectives To examine relationships between a comprehensive set of linguistic features derived from logical memory delayed recall (LMDR) responses and amyloid status as well as regional tau burden among cognitively normal participants. Design In this cross-sectional study, we evaluated associations between two types of AD biomarkers (amyloid status and regional tau burden) and linguistic features using covariate-adjusted regression models tailored to each feature’s distribution (e.g., linear, negative binomial, zero-inflated beta, and gamma regression). We then used LASSO regression and the Boruta feature-selection algorithm to identify features most informative of biomarker status. Setting A community-based subcohort from the Framingham Heart Study. Participants A total of 238 cognitively unimpaired participants who completed the LMDR test within 1 year before or after tau positron emission tomography (PET) were included. Of these, 228 also underwent amyloid PET and 227 had LMDR voice recordings available. Measurements Verbal responses to the LMDR test were digitally recorded and manually transcribed. We extracted 54 linguistic and paralinguistic features from transcribed responses using natural language processing, capturing lexical diversity and density, syntactic complexity, semantic/information density, and speech fluency and rate. Tau SUVR values were quantified from PET imaging in 5 brain regions (amygdala, entorhinal cortex, inferior temporal [IT], inferior parietal [IP], and precuneus). Amyloid status was classified as Aβ− or Aβ+ using Gaussian mixture modeling of PiB PET DVR values. LMDR total score and participant demographics (age, sex, education) were included as covariates. Results After adjusting for age, sex, education, amyloid status, and LMDR score, 23 linguistic features were associated with tau burden in at least one brain region. For example, reduced semantic richness, captured by lower uniqueIdeaDensity, was associated with higher tau burden across multiple regions, including the amygdala (average marginal effect (AME)=-0.007, 95% CI: (-0.013, -0.001)), entorhinal (AME=-0.007, 95% CI: (-0.013, -0.002)), IT (AME=-0.008, 95% CI: (-0.014, -0.002)), and precuneus (AME=-0.006, 95% CI: (-0.012, -1.51e^-5)). More between-utterance pauses were associated with higher tau burden in the entorhinal (AME=0.259, 95% CI: (0.026, 0.493)), IT (AME=0.272, 95% CI: (0.025, 0.518)), and IP (AME=0.404, 95% CI: (0.160, 0.648)) regions. No linguistic features were associated with amyloid status, and LMDR total score was not associated with either tau burden or amyloid status. LASSO regression selected 3–7 features per brain region, capturing multiple aspects of language use (e.g., speech disfluency and semantic density) associated with tau burden. The nonlinear Boruta algorithm selected a larger set of biomarker-informative features, including predictors of amyloid positivity, and identified additional syntactic measures (e.g., Yngve score, mean dependency distance) as important predictors of regional tau burden. Conclusions Linguistic features derived from neuropsychological (NP) tests capture subtle language changes associated with regional tau pathology in cognitively normal individuals, beyond conventional NP test scores. These findings highlight the potential of language-based measures as non-invasive digital biomarkers for early AD risk stratification and for supporting clinical decision-making about biomarker testing for possible preclinical AD.