Brain Health Prediction in Older Adults via Digital Speech Biomarkers

Background: 

Increased life expectancy worldwide has been accompanied by a rising prevalence of age-related neurocognitive conditions. These occur along a continuum, encompassing subjective cognitive decline (SCD), mild cognitive impairment (MCI), and Alzheimer’s disease dementia (ADD), and are commonly evaluated using cognitive and neuroimaging measures. These approaches are costly and require specialized personnel or resources that are scarce in many regions. Digital speech biomarkers (DSBs) offer an affordable, scalable alternative. To date, however, no study has leveraged DSBs to predict neuropsychological and brain measures across this continuum. Here we tackle this gap focusing on Spanish-speaking Latinos.

Methods: 

Participants spanning the SCD–MCI–ADD continuum (n = 150; 17 healthy controls, 55 with SCD, 57 with MCI, 21 with ADD) completed brief phonemic and semantic fluency tasks, standardized cognitive and executive function testing (Addenbrooke’s Cognitive Examination–III [ACE-III], INECO Frontal Screening [IFS]), and MRI protocols. DSBs features were extracted from the fluency tasks using the TELL app; cognitive features were derived from ACE-III and IFS scores; and brain features were computed from the volumes of dementia-sensitive regions. Machine learning models were trained using DSBs and brain features separately to predict ACE-III and IFS scores, and using DSBs and cognitive features separately to predict hippocampal volume and an ADD-related atrophy mask. Model performance was quantified by age-adjusted partial Pearson correlations between predicted and observed values. To investigate neural correlates, multiple regression analyses were conducted between DSBs and gray matter volume.

Results: 

DSBs features outperformed brain features in predicting ACE-III (r = .67 vs. .23) and IFS scores (r = .55 vs. .31). They also outperformed cognitive features in predicting ADD mask volume (r = .38 vs. .27), but not in predicting hippocampal volume (r = .37 vs. .42). Word frequency was negatively correlated with prefrontal cortex volume.

Conclusion: 

DSBs capture both cognitive performance and MRI-derived measures of brain health across the SCD–MCI–ADD continuum, surpassing standard cognitive and imaging measures. By offering a scalable, objective, time-efficient, and affordable alternative to gold-standard assessments, they support more equitable approaches to dementia detection and monitoring worldwide.