Speech-based estimates of aging could differ between healthy controls and patients with several dementia phenotypes.
Investigators analyzed 2,928 native Spanish-speaking participants recruited through the Multi-Partner Consortium to Expand Dementia Research in Latin America, including 1,504 healthy controls, 24 patients with mild cognitive impairment (MCI), 1,068 patients with Alzheimer's disease (AD), 255 patients with non–language-dominant frontotemporal dementia (nldFTD), and 77 patients with language-dominant frontotemporal dementia (ldFTD). The investigators extracted acoustic and linguistic features spanning speech timing, pitch, semantic characteristics, lexical features, emotional content, and verbosity among the participants asked to complete 7 standardized Spanish-language speech tasks.
Using supervised models, they estimated chronological age from the speech features and calculated a speech age gap (SAG), defined as predicted speech age minus chronological age. Positive SAGs represented older-appearing speech profiles, whereas negative values represented younger or preserved profiles. The investigators assessed differences the associations between SAGs and cognition, plasma phosphorylated tau 217 (p-Tau217), social exposome, neuroimaging-derived brain-age clocks, and epigenetic aging clocks.
The speech model explained 44% of the variance in chronological age. SAGs showed a graded pattern across diagnostic groups, with healthy controls having lower gaps compared with those with dementia. Among dementia phenotypes, patients with ldFTD followed by those with nldFTD had larger SAGs compared with patients with AD. The diagnostic pattern remained consistent in analyses performed without data augmentation.
Larger SAGs were associated with poorer performance across clinical and cognitive measures. Patients with AD and nldFTD showed consistent associations between larger SAGs and poorer performance across all assessed clinical and cognitive domains. Associations were stronger for linguistic than nonlinguistic performance.
The speech-derived measure also converged with independent markers of neurodegeneration and social exposures. Across the participants with available biomarker data, larger SAGs were associated with higher plasma p-Tau217. Within diagnostic groups, the association was observed in patients with AD, but not in healthy controls or patients with nldFTD or ldFTD. Larger SAGs were associated with a more adverse social exposome, with within-group associations observed among healthy controls and patients with AD.
The associations were strongest for neuroimaging-derived measures of biological aging. SAGs correlated with structural, functional, and combined brain-age gaps. Positive associations between speech and brain-age gaps were observed within healthy controls and patients with AD, nldFTD, and ldFTD.
Associations with epigenetic aging were smaller. SAGs correlated with the Hannum, Retroclock, and OMICmAge epigenetic clocks. Healthy controls and patients with AD showed associations across all 3 epigenetic clocks, whereas patients with ldFTD showed associations with Retroclock and OMICmAge. Patients with MCI could not be included in within-group brain or epigenetic clock analyses because no patients had complete speech data and corresponding aging-clock measures.
Sensitivity analyses indicated that the principal diagnostic differences persisted after accounting for age, sex, country, education, comorbidities, and number of languages spoken. Propensity matching for age, sex, and education preserved the principal SAG differences. The investigators further reported that the composite SAG measure provided stronger discrimination across the clinical contrasts compared with individual linguistic, semantic-memory, or acoustic domains.
The study had several limitations. Its primarily cross-sectional design precluded causal inference and did not establish whether SAGs reflected changes in aging within patients over time or predicted future clinical conversion. Multicenter differences in recording conditions, microphones, and speech prompts may have introduced residual variability, and automated speech-feature extraction could incorporate dialectal or cultural biases. Some diagnostic groups and several biomarker subsets were small, and biomarker and social exposome data were incomplete for some participants. The cohort was restricted to Latin American populations completing structured Spanish-language speech tasks, limiting generalizability to other languages, cultures, and naturalistic conversational settings.
“Speech clocks afford a shortcut to multimodal aging in health and disease,” wrote lead study author Hernan Hernandez, of the Latin American Brain Health Institute at Universidad Adolfo Ibañez in Chile, and colleagues.
Full disclosures of the study authors can be found in the study.
Source: Science Advances
