Bridging Language Markers and Pathology: Correlations Between Digital Speech Measures and Surrogate CSF Biomarkers in Alzheimer’s Disease

Yijiang Pang, Liu Chen, Hiroko H. Dodge, Jiayu ZhoubioRxiv 2025 · DOI

TL;DR

This study bridges the gap between digital language markers and Alzheimer's disease (AD) pathology by using machine learning to predict cerebrospinal fluid (CSF) biomarkers from clinical data. They found that language features, like syntactic complexity and response length, correlate with predicted AD pathology markers, suggesting conversational analysis could be a scalable, non-invasive screening tool.

What Problem Does This Paper Address?

Alzheimer's disease (AD) is characterized by amyloid-beta (Aβ) plaques and tau tangles, which can be measured by CSF biomarkers. However, obtaining CSF samples via lumbar puncture is invasive and costly, limiting widespread use. While digital speech and language markers show promise for detecting early cognitive impairment, their direct link to AD pathology (not just clinical status) has been unclear due to the difficulty of collecting simultaneous speech and fluid biomarker data. This study aims to establish this link by developing a method to infer AD pathology markers from language data.

Methods

The core innovation here is the development of a "surrogate biomarker modeling strategy." * Training Biomarker Predictors: They used the extensive National Alzheimer's Coordinating Center (NACC) database, which contains clinical data and CSF biomarker measurements, to train machine learning models (specifically LightGBM) to predict Aβ42, P-tau181, and T-tau levels from 168 common clinical variables. * Applying to a New Cohort: These trained models were then applied to participants from the I-CONECT study, who had rich conversational speech data but no CSF biomarkers. By inputting their clinical data into the NACC-trained models, they generated "surrogate" CSF biomarker estimates for the I-CONECT participants. * Correlation Analysis: Finally, they analyzed the correlations between these surrogate CSF biomarker estimates and various linguistic features extracted from the I-CONECT participants' conversations. * Distribution Shift Mitigation: To address potential differences between the NACC and I-CONECT datasets, they used Principal Component Analysis (PCA) to identify and remove outliers from the I-CONECT data before prediction, ensuring the models were applied to data within a more reliable range.

Datasets

  • NACC (National Alzheimer's Coordinating Center) Database: This dataset served as the primary source for training the machine learning models to predict CSF biomarkers. It contains extensive clinical evaluations and biomarker data from Alzheimer's Disease Research Centers (ADRCs) across the US.
  • I-CONECT (Internet Based Conversational Engagement Clinical Trial) Study: This dataset provided the conversational speech data from socially isolated older adults with normal cognition or mild cognitive impairment. This cohort was used for extracting language features and correlating them with the surrogate biomarkers.

Key Insights and Ideas

  • Bridging the Gap: The study successfully bridges the gap between traditional, invasive AD biomarkers (CSF) and non-invasive digital markers (speech/language) by creating a computational link.
  • Conversational Data as a Proxy: Naturalistic conversational speech, collected in a real-world social context, contains signals related to underlying AD pathology, not just clinical symptoms.
  • Scalability for Screening: This approach offers a path towards a more scalable and accessible screening tool for AD, potentially identifying individuals at risk much earlier than current methods.
  • Language as a Window into Pathology: The correlations suggest that specific language features are not just proxies for cognitive status but may reflect the biological processes of AD itself.

Key Results

  • Predictive Models: The LightGBM models achieved reasonable accuracy in predicting CSF biomarker values, with cross-validated R² values ranging from approximately 0.54 to 0.87 for the different biomarkers.
  • Language Correlations:
    • Lower predicted Aβ42 (indicating higher amyloid burden) was significantly correlated with shorter average response lengths (p ≈ +0.3) and reduced syntactic complexity (p ≈ +0.3).
  • Higher surrogate tau levels were associated with lower syntactic complexity (p ≈ -0.3).
  • Feature Importance: Cognitive scores were among the strongest predictors of biomarker levels in the models, aligning with established knowledge about AD.

Limitations and Future Work

  • Surrogate Biomarker Imperfection: The surrogate biomarker estimates are acknowledged as imperfect proxies, and external validation with actual CSF or blood biomarkers in the same individuals is needed.
  • Domain Shift: Differences between the NACC and I-CONECT cohorts (e.g., social isolation in I-CONECT) present a domain shift challenge, though steps were taken to mitigate this. More sophisticated domain adaptation techniques are suggested.
  • Cross-Sectional Data: The current analysis is cross-sectional, limiting the ability to determine causal directionality. Longitudinal studies are proposed to clarify whether pathology drives language changes or vice-versa.
  • Confounding Factors: While regression models adjust for many factors, residual confounding from variables like education or hearing loss is possible. Future work will explore multivariate models including these covariates.
  • Future Directions: The authors envision a tiered screening system where conversational analysis combined with predictive models could generate an "AD biomarker risk score," guiding individuals towards confirmatory testing or early intervention. Longitudinal analyses are also planned to track changes and explore personalized intervention strategies.

Summary updated

Abstract

This study demonstrates novel links between language markers and fluid biomarkers, highlighting conversational language as a potential accessible, non-invasive approach for early detection and monitoring of Alzheimer's pathology.
BibTeX
@misc{pang2025bridging,
  author = {Yijiang Pang and Liu Chen and Hiroko H. Dodge and Jiayu Zhou},
  title = {{Bridging Language Markers and Pathology: Correlations Between Digital Speech Measures and Surrogate CSF Biomarkers in Alzheimer’s Disease}},
  year = {2025},
  doi = {10.1101/2025.11.05.686472},
  howpublished = {bioRxiv},
}
Figure