Spectral signatures of Alzheimer’s disease: Transformer-based analysis of speech patterns toward explainable detection and monitoring

The early and accurate detection of neurocognitive decline in adults promises to broaden the eligibility scope of potential research subjects for more robust longitudinal monitoring of the effect of novel therapies and therapeutics that seek to manage, slow, or halt the underlying mechanism(s) of neurodegenerative conditions such as Alzheimer's Disease. As earlier stages of its progression become more readily available for study, it is expected that a more holistic understanding of Alzheimer's symptomatologies and onset conditions will be revealed which could in turn lead to a cure. Acoustic speech signals have been suggested as a lightweight and cost-effective proxy for standard neurocognitive screening tests which are built, at least in part, on measures of linguistic production. This research aims to model differences in spectrographic representations of speech produced by individuals diagnosed with Alzheimer's Disease and Mild Cognitive Impairment as compared with healthy aging adults using statistical and machine learning techniques that simultaneously provide a much needed update to model comparability while also furnishing clinicians with insights into the \textit{why} behind neural \say{black box} decisions. The audio recordings used throughout this study were provided mainly through DementiaBank, comprising extemporaneous speech elicitation interviews. These were used to fit an audio spectrogram transformer that produced attention maps highlighting time-frequency regions of interest en route to discriminating between study populations. The final configurations did not outperform the state-of-the-art community baselines in terms of accuracy and error rates. However, the attention map analysis provides a novel pathway toward greater collaboration between speech-language pathologists and machine learning practitioners while the introduction of statistical calibration to this interdisciplinary research space stands to facilitate a more grounded discourse on comparing novel featurization and modeling approaches. These results lay a foundation upon which to build more explanatory models that can overcome the apparent trade-off between performance and interpretability toward high quality, language agnostic early detection of elements of neurocognitive decline.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Linguistics

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Adams, Chase

Contributors dc:contributor

• Tang, Yan
• Mudar, Raksha
• Schwartz, Lane
• Shih, Chilin
• Shosted, Ryan

• Alzheimer's disease
• attention map
• audio spectrogram
• calibration
• mild cognitive impairment
• neurocognition
• transformer neural network

Statement dc:rights

• Copyright 2025 Chase Adams

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/129471

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F129471