Poster Summary

Changes to speech patterns have been identified as early signs of Alzheimer’s Disease (AD) and have been shown to progress with disease, but it is not known if changes to speech that occur prior to any detectable cognitive decline can be used to identify risk of AD. We used data-driven approaches to identify speech phenotypes in a sample of cognitively healthy participants at risk of developing AD.

We analysed baseline Clinical Dementia Rating (CDR) interview recordings from 114 participants, aged 59-76, who were cognitively healthy but had risk factors for developing AD. CDR recordings were segmented, diarized and transcribed and 8 categories of speech features were extracted using the Winterlight speech analysis platform. We performed a dimensionality reduction analysis (PCA) within each feature category and extracted the first two components. Blinded k-means cluster analysis was performed to determine if participants clustered into subgroups based on speech features.

Silhouette analysis yielded two clusters of participants who differed on the timing and acoustic feature categories in the “address repeat” and “recent experience” sections.  Cluster analysis indicated that at baseline, before any sign of cognitive decline, participants could be distinguished into two groups based on timing and acoustic parameters of speech. These clusters did not differ on conventional clinical endpoint scores (e.g. RBANS and MMSE) indicating that speech measures may detect subgroups at preclinical stages of AD that do not differ on cognitive assessments.

The project demonstrates how data-driven methods can identify speech phenotypes from naturalistic, passively collected speech recordings.

View the poster