A Multidisciplinary Investigation of Conversation and Disfluencies in Cognitive Decline

Cognitive Decline (CD) encompasses a spectrum of conditions affecting millions globally, manifesting in cognitive impairments such as memory and language deficits. Neurodegenerative Dementias (NDs), including Alzheimer's Dementia, represent a group of degenerative disorders contributing to progressive CD. The early stages of CD often exhibit language disturbances, and research indicates that early diagnosis can improve patient outcomes. Speech has emerged as a prominent, non-invasive biomarker for CD assessment, offering potential insights into disease progression. Studies investigating how speech is affected by CD have frequently reported that, as cognition decreases, the presence of disfluencies such as unfilled pauses increases. This thesis explores the diagnostic utility of disfluency analysis, as well as investigating which tasks may elicit the most useful speech for analysis.

In CD detection, advancements in machine learning have led to the development of Automatic Cognitive Decline Classification (ACDC) systems, which demonstrate remarkable accuracy in distinguishing dementia patients from healthy controls based on speech samples. However, ACDC methodologies often struggle to generalise across diverse demographics and lack transparency in their classification rationale. This thesis presents evidence that integrating disfluency features into ACDC systems enhances classification accuracy and addresses issues of generalisation and transparency.

Additionally, Conversation Analysis (CA) has been employed to develop conversational profiles that could assist doctors in differentiating between patients with ND and those with Functional Memory Disorder (FMD), a non-neurodegenerative psychological condition. This thesis further investigates whether CA can be utilised to create conversational profiles that help differentiate between ND, FMD, and Mild Cognitive Impairment, an early stage of CD.