# Testing speech biomarkers against cognitive and neural signatures of Alzheimer's dementia

## TL;DR

This study shows that analyzing speech patterns can detect Alzheimer's disease as effectively as traditional methods, using just two minutes of audio.

## Contribution

The study demonstrates that speech biomarkers perform comparably to standard cognitive and neural measures for Alzheimer's detection.

## Key findings

- Speech biomarkers achieved an AUC of 0.88 and accuracy of 0.81 for Alzheimer's detection.
- Top speech features included word concreteness and semantic variability, correlating with brain volume changes.
- Digital speech analysis matched the performance of gold-standard cognitive and imaging assessments.

## Abstract

Automated analysis of word properties (WP) and speech timing (ST) offers a cutting‐edge digital method for identifying scalable markers of Alzheimer's disease dementia (ADD). Quantification of lexical features during speech can not only detect ADD cases but also predict cognitive performance and associated anatomical‐functional brain patterns. However, the clinical value of WP/ST markers remains uncertain, as no study has yet compared their discriminatory capacity to standard cognitive and neural measures, casting doubt on their utility for under‐resourced settings that lack such conventional tools.

We recruited 33 ADD patients and 33 healthy controls from a carefully characterized cohort. Each participant completed two 1‐minute verbal fluency tests, cognitive evaluations (addressing general cognitive skills, attention, set‐shifting, and working memory), and MRI scans. Separate machine learning classifiers were trained using (i) WP/ST features from the fluency tasks (obtained through the Toolkit to Examine Lifelike Language [TELL] app), (ii) scores from the cognitive tests, and (iii) volumetric features from the imaging protocol. The best‐performing model for each feature set was then evaluated based on mean AUC, accuracy, and 95% confidence intervals. Additionally, we identified the top discriminatory WP/ST markers through feature importance analysis and explored their correlation with gray matter volume.

Classification outcomes revealed comparable performance between WP/ST features (AUC = 0.88; accuracy = 0.81) and both cognitive (AUC = 0.86; accuracy = 0.79) and imaging (AUC = 0.92; accuracy = 0.88) features, with overlapping confidence intervals suggesting similar discriminability. The top five WP/ST features were word concreteness, semantic variability and frequency for the semantic fluency task, and semantic variability and granularity for the phonemic task. Across all participants, word frequency showed a negative correlation with the volume of right prefrontal areas involved in executive processing.

Using only two minutes of audio recordings, our digital speech analysis pipeline can identify ADD with performance comparable to gold‐standard measures, driven by features linked to brain regions implicated in cognitive symptoms typical of ADD. Overall, WP/ST analyses seem non‐inferior to standard diagnostic measures, highlighting their potential as a scalable and low‐cost tool for dementia assessment.