Dementia is a gradual cognitive decline that typically occurs as a consequence of neurodegenerative disease, and can result in language deficits (i.e., aphasia). I show that linguistic features automatically extracted from the connected speech samples of individuals with dementia can both differentiate these individuals from healthy controls and contribute to our knowledge of the nature of language impairment in dementia. As a secondary goal, I address the challenges of a fully automated processing pipeline.

I focus on a dementia syndrome known as primary progressive aphasia (PPA), in which language abilities are specifically impaired. I begin by automatically extracting linguistic information from transcripts of PPA speech, training machine learning classifiers to differentiate between the different variants of PPA relative to healthy controls, and interpreting the selected features in the context of the PPA literature. While traditional measures of syntactic complexity do not distinguish between the groups, the inclusion of parse-based syntactic features ultimately leads to accuracies of over 90% in three classification tasks.

Having shown that the extracted features can differentiate the groups, I examine how these features degrade as a result of the processing steps in a fully automated pipeline, including automatic speech recognition (ASR) and sentence segmentation. The classifiers still achieve positive results, although the degraded feature accuracy may be of concern in biomedical applications.

I then explore a question of some debate in the literature:​ Is there a difference between agrammatism in PPA and agrammatism in post-stroke aphasia? Above-baseline classification results suggest that there are indeed differences between these two impairments.

Having validated the methodology on PPA, I conclude by examining whether a similar analysis will detect Alzheimer's disease from speech samples, even though language impairment is not the primary symptom of the disease. By including additional features to measure the information content of the narrative, classification accuracies of up to 81% are achieved. I repeat the classification experiment using ASR transcripts, and find that many of the relevant features are still significantly different between the groups, suggesting that a fully automated analysis may be possible as ASR improves.