Predicting brain health in older adults through automated language measures

Gonzalo Pérez, Iván Caro, Joaquín Valdez Bisé, Franco Javier Ferrante, Lara Gauder, Alejandro Sosa Welford, Joaquín Ponferrada, Luciana Ferrer, Agustin Ibanez, Andrea Z. Slachevsky, Adolfo M. GarcíaAlzheimer 2025 · DOI

TL;DR

This paper introduces an automated, language-based approach to assess brain health in older adults, offering a scalable and affordable alternative to traditional cognitive and neuroimaging assessments. By analyzing speech patterns from brief fluency tasks, the method accurately predicts cognitive performance and brain structure, paving the way for more equitable global access to neurocognitive screening.

What Problem Does This Paper Address?

As global life expectancy rises, so does the prevalence of age-related neurocognitive challenges, ranging from subjective cognitive decline (SCD) to mild cognitive impairment (MCI) and Alzheimer's disease dementia (ADD). While current diagnostic methods like cognitive tests and neuroimaging (e.g., MRI) are effective, they are often expensive, require specialized personnel, and can involve significant waiting times. This creates a barrier to early detection and intervention, particularly in underserved regions. The paper aims to bridge this gap by developing an automated, accessible, and cost-effective method for assessing neurocognitive health.

Methods

The core of this research lies in an automated speech and language analysis (ASLA) pipeline. Participants completed brief (1-minute) phonemic and semantic fluency tasks. The audio recordings from these tasks were then analyzed using the TELL app to extract various features, including word-property (e.g., word frequency, semantic relatedness) and speech-timing characteristics. Machine learning models were trained using these extracted features to predict cognitive test scores (ACE-III, MoCA, IFS) and MRI-derived brain measures (total gray matter volume, white matter volume, hippocampal gray matter volume, and ADD-sensitive regions). The models employed a multivariate approach, and the best predictors were identified using confidence intervals of R² scores. Pearson's partial correlations were used to assess the relationship between actual and predicted values, controlling for age, sex, and education (and MoCA scores for brain measures). Analyses were performed on individual fluency tasks, their combination, and their average.

Datasets

The study utilized data from 150 Chilean individuals. This cohort was stratified into four groups:

  • 17 healthy controls
  • 55 individuals with subjective cognitive decline (SCD)
  • 57 individuals with mild cognitive impairment (MCI)
  • 21 individuals with Alzheimer's disease dementia (ADD)

The paper does not explicitly state the use of pre-existing public datasets for training or evaluation, suggesting the data was collected specifically for this study.

Key Insights and Ideas

  • Language as a Window to Brain Health: The study powerfully demonstrates that subtle patterns in spoken language, captured through brief fluency tasks, can serve as a sensitive indicator of underlying neurocognitive status and brain structure.
  • Scalability and Accessibility: The ASLA approach offers a significant advantage in terms of scalability and accessibility. Unlike resource-intensive traditional methods, automated language analysis can be deployed widely and affordably, democratizing access to neurocognitive screening.
  • Examiner Independence: The automated nature of the pipeline removes the variability and potential bias associated with human examiners, leading to more objective and reproducible assessments.
  • Multivariate Feature Integration: The success of combining various word-property and speech-timing features highlights the power of a holistic linguistic analysis rather than relying on single linguistic markers.

Key Results

The automated language analysis pipeline demonstrated significant predictive power for several key measures:

  • Cognitive Performance:
    • ACE-III scores: Significant partial Pearson correlation of r = .55 (p < .001)
    • MoCA scores: Significant partial Pearson correlation of r = .39 (p < .001)
    • IFS scores: Significant partial Pearson correlation of r = .31 (p = .001)
  • Brain Structure:
    • Normalized gray matter (GM) volume: Significant partial Pearson correlation of r = .34 (p = .006)

These results were achieved using word-property features derived from combined fluency tasks.

Limitations and Future Work

The authors acknowledge that this study was conducted on a specific population (Chilean individuals) and that further validation is needed across diverse linguistic and cultural backgrounds. While the results are promising, the paper does not delve deeply into the specific machine learning algorithms used beyond mentioning "machine learning models." Future work could explore:

  • Cross-cultural validation: Replicating the study in different Spanish-speaking populations and other languages.
  • Longitudinal studies: Investigating the predictive capabilities of ASLA for tracking cognitive decline over time.
  • Algorithm refinement: Exploring a wider range of machine learning techniques and feature engineering approaches.
  • Clinical integration: Developing user-friendly tools for clinicians and researchers to easily implement this ASLA approach.

Summary updated

Abstract

A fully automated, multivariate pipeline captures cognitive and MRI measures of brain health based on brief fluency tasks. Unlike gold-standard tests, this approach is examiner-independent, objective, time-efficient, and affordable. This approach represents a scalable resource to favor equity in the global fight against dementia and cognitive decline.
BibTeX
@inproceedings{perez2025predicting,
  author = {Gonzalo Pérez and Iván Caro and Joaquín Valdez Bisé and Franco Javier Ferrante and Lara Gauder and Alejandro Sosa Welford and Joaquín Ponferrada and Luciana Ferrer and Agustin Ibanez and Andrea Z. Slachevsky and Adolfo M. García},
  title = {{Predicting brain health in older adults through automated language measures}},
  booktitle = {Alzheimer},
  year = {2025},
  doi = {10.1002/alz70857\_102831},
}
Figure