Toward Addressing the Low Resource Task of Dementia Detection through Spontaneous Speech

As populations grow older worldwide, the number of people with Alzheimer’s disease (AD) and related dementia is also on the rise. Significant changes to speech and language use caused by dementia occur early in disease progression, and automatically detecting these changes from spontaneous speech may pave the way toward more cost-effective, non-invasive approaches for analyzing and understanding early signs of dementia. However, effective training of such models is challenging due to small dataset sizes, imbalanced datasets that are fragmented towards different tasks and domains, and high transcription costs. In this dissertation, we study the problem of dementia detection through spontaneous speech from three aspects: 1) we propose task- and domain-adaptive methods to overcome issues associated with low resource availability for classifying dementia, 2) we model fine-grained cognitive health status from a battery of tests using spontaneous speech data, and 3) we introduce and analyze a new spoken language corpus designed specifically to facilitate early AD detection. During the course of this work, we have collected dialog act (DA) labels to foster the study of speech patterns in AD and empirically validated the efficacy of novel DA-based interaction features for task-agnostic dementia detection. Moreover, we established strong baselines for predicting Mini Mental State Examination (MMSE) scores using feature-based machine learning models and pre-trained language models, finding that verbal and non-verbal speech and language biomarkers are strong predictors of MMSE scores. We then present a systematic analysis of the use of feature-based domain adaptation and domain-adaptive prompt finetuning of pretrained language models on heterogeneous spoken language data to evaluate the generalizability across diverse datasets for dementia detection. We empirically prove that adapted models exhibit better performance across conversational and task-oriented datasets and we identify robust linguistic and acoustic features for dementia prediction across domains. Finally, we present a novel spoken language corpus, SLaCAD, with paired spontaneous speech and clinical biomarkers indicating different stages of dementia, with a focus on preclinical AD. We establish exciting preliminary benchmarks on this dataset for detecting early AD using speech and language biomarkers, and also perform feature understanding experiments, using statistical tests to discern the importance of features and their effects across diagnostic and demographic categories.