Speech Clock Tested in Spanish-Speaking Latin America

A Potential Speech-Based Aging MarkerA study published in Science Advances suggests that speech may contain measurable signals of aging. Researchers led by principal investigator Hernán Hernández, PhD, a professor at Universidad Adolfo Ibañez in Santiago, Chile, developed a machine-learning “speech clock” that estimated participants’ ages using acoustic and linguistic features.Hernán Hernández, PhDThe gap between estimated and chronological age was greater among people with Alzheimer’s disease or frontotemporal dementia than among cognitively healthy adults. The gap was also associated with plasma phosphorylated tau 217 levels, cognitive performance, social adversity, brain age, and epigenetic age.The significance of this finding lies in its potential scalability, especially in regions such as Latin America. Currently available aging clocks — such as brain, epigenetic, or biobehavioral ones — rely on MRI scans, blood samples, or battery tests administered by specialized personnel, making them difficult to access in low-resource settings. In contrast, the authors state that speech, which can be recorded and transcribed automatically, remotely, without specialized personnel, and at low cost, could be a candidate for a scalable and culturally adaptable biomarker for research on aging and dementia in underrepresented regions, with future applications in large-scale surveillance, early detection, and prevention, as it is a voice-derived measure that converges with molecular, cerebral, cognitive, and social indicators.The study arose from the need for objective, automated, affordable, and noninvasive aging markers that can be deployed in low-income regions, as well as from evidence that speech is sensitive to the cognitive and motor changes associated with aging and dementia.“In our group, we’ve been looking for algorithms that attempt to predict people’s age — which we call ‘clocks’ — that calculate the difference between chronological age and the age predicted by the clock; this gap allows us to determine whether the clock distinguishes between control group and patients, for example, those with Alzheimer’s disease, and whether it’s also related to variables such as inequalities or exposomes. But since we’re in Latin America, we’re looking for predictors or variables that are more cost-effective than MRI scans, which are the gold-standard ‘clocks,’” Hernández told Medscape’s Spanish edition. “Recording audio is an inexpensive biomarker that can be done remotely; it’s environmentally friendly and can easily be performed without requiring specialized testing.”How the Speech Clock WorksThis was a cross-sectional, multicenter study involving 2928 native Spanish speakers with a mean age of 65.2 years, 67% of whom were female, recruited through the multi-partner consortium to expand dementia research in Latin America (ReDLat). Participants from Colombia, Peru, Mexico, Chile, and Argentina included 1504 cognitively healthy control participants, 24 people with mild cognitive impairment, 1068 with Alzheimer’s disease, 255 with nonlanguage-predominant frontotemporal dementia, and 77 with language-dominant frontotemporal dementia.Neuropsychologists recorded each participant performing seven standardized tasks, including narrating the video description tasks, two phonological-fluency tests, two semantic fluency tests, and immediate and delayed retelling of a story. An automated preprocessing and transcription workflow extracted more than 700 features per recording, summarized into seven categories: speech timing, pitch, concreteness, granularity, lexical features, emotional content, and verbosity. The researchers used these features to train a supervised regression model to estimate participants’ chronological ages. The model then calculated the speech age gap (SAG) for each participant, defined as the predicted age minus chronological age; a positive gap indicated a speech profile that appeared older than expected. The researchers examined associations between SAGs and plasma phosphorylated tau 217 levels, a multimodal social exposome measure, brain-age gaps derived from structural and functional MRI, and epigenetic age estimates based on DNA methylation in peripheral whole blood.Until now, no study had examined speech age gaps or differences between a person’s predicted speech age and chronological age. The research team found that slower processing speed results in longer pauses and syllables and slower speech, while affective flattening is reflected in tone and emotional content, and changes in semantic memory are reflected in the conciseness and precision of vocabulary. It was also observed that verbosity and speech temporality were the categories most strongly associated with preserved aging, while tone and emotional content were most strongly associated with accelerated aging. The control group had a smaller speech age gap than all patient groups, with the highest value observed in language-predominant frontotemporal dementia, followed by nonlanguage-dominant frontotemporal dementiaand Alzheimer’s disease.Findings Across Diagnostic Groups“The speech clock we developed allowed us to clearly distinguish, in a stepwise manner, between controls and individuals with a cognitive condition that accelerates their aging; furthermore, this was associated with social exposomes, such that the more adverse a person’s life was on an individual level, the greater the acceleration of aging,” commented Hernández, adding, “We were surprised that our speech age gap was more strongly correlated with brain clocks — such as those measured by MRI, which are the established standard — than with epigenetic aging clocks like phosphorylated tau 217, showing that the speech clock could have the potential to be used as a very low-cost diagnostic tool.”The model predicted chronological age with moderate performance (coefficient of determination [R2] = 0.44; r = 0.67; mean absolute error, 9.04 years). A larger speech age gap was associated with higher concentrations of phosphorylated tau 217 (r = 0.21; P = 3.3 × 10-6), a relationship that was significant by group only in Alzheimer’s disease (r = 0.18; P = .01) and with worse clinical and cognitive performance (r = -0.21 to -0.38), consistently across all domains in Alzheimer’s disease and in nonlanguage-dominant frontotemporal dementia, particularly in memory (r = -0.34 and -0.42, respectively).The association was stronger with linguistic tasks (r = -0.38) than with nonlinguistic tasks (r = -0.31) but was present in both. A more adverse social exposome (education level, food insecurity, financial status, access to health services, and early-life experiences, among other components) was associated with a larger speech age gap (r = -0.22 with the exposome score, where higher values indicate more favorable conditions), with statistical significance in the control group and in Alzheimer’s disease.The speech age gap correlated with structural (r = 0.42), functional (r = 0.49), and combined (r = 0.53) brain age gaps in both the control group and the three dementia groups, as well as with epigenetic age acceleration according to the Hannum, Retroclock, and OMICmAge clocks (r = 0.20-0.21), consistently in the control group and in Alzheimer’s disease, and in language-dominant frontotemporal dementia, only with the latter two. Differences between diagnoses persisted after adjusting for sex, country, education level, age, comorbidities, and number of languages spoken.Expert Perspectives and LimitationsAdolfo M. García, PhD“Our study reinforces the importance of speech and language for understanding brain health,” Adolfo M. García, PhD, the study’s corresponding author and a specialist in language neuroscience at the University of San Andrés in Buenos Aires, Argentina, told Medscape’s Spanish edition. “In previous studies, we have shown that digital speech markers can detect neurodegenerative diseases, assess their severity, and predict cognitive abilities, brain alterations, and even plasma biomarkers, with accuracy comparable to that of standard clinical and neurobiological measures. This research expands the clinical potential of this approach, validating language as a scalable window into brain aging.”Luana Sandoval CastilloLuana Sandoval Castillo, a physician specializing in orthogeriatrics and cardiology at Bispebjerg Hospital in Copenhagen, Denmark, who did not participate in the study, noted that “this is an extremely innovative and methodologically interesting finding, but it is still in the discovery phase; it is an initial validation that cannot yet be used as a diagnostic tool for patients.” “It will be interesting to see the follow-up — the longitudinal study of the speech age gap — since it is related to biological markers, which could eventually lead to a language-based screening test years before the onset of a disease and provide a diagnostic tool in hard-to-reach regions,” she added, noting that it is significant that the study was conducted in Spanish; however, “it would also be interesting to examine the effectiveness of the speech age gap in people who speak other languages and even to investigate the protective cognitive factor associated with multilingualism.”The next steps are shaped by the limitations acknowledged by the authors themselves. The cross-sectional design prevents inferring causality, describing individual trajectories, or predicting clinical conversion; thus, longitudinal studies are required to determine the prognostic value and predictive clinical utility of the speech age gap. “For it to become a potential diagnostic tool, we will use the same cohort to conduct a longitudinal study and verify whether, over time, the same participants can distinguish aging patterns among diagnostic groups, controls, and at the individual level,” commented Hernández.It will also be necessary to validate it in other languages and cultures, ideally with language-specific normalization or multilingual training models, and in conversational contexts that are more natural than structured tasks. Furthermore, biomarker and exposome data were not available for all participants, which raises the possibility of selection bias. “A major challenge is coordinating so many people involved in the study and processing so many recordings and data sets through the algorithms,” added Hernández, “but these are the next steps through ReDLat.”Potential Future ApplicationsIf the findings are confirmed, the authors believe that speech clocks could be integrated into global aging research, clinical workflows, and public health initiatives aimed at preventing dementia and promoting equity in brain health; the model’s code is publicly available.Hernández, García, and Sandoval declared having no relevant financial conflicts of interest.This story was translated from Medscape’s Spanish edition.