Automatic screening of dementia and age-associated cognitive decline from conversational speech

Alzheimer’s disease (AD) and Age-Associated Cognitive Decline (AACD) pose significant challenges to aging populations worldwide, making early detection crucial for timely intervention. Traditional assessments are effective but resource-intensive, limiting their scalability for widespread and casual screening. Speech and language have emerged as promising biomarkers for cognitive decline, as they reflect complex cognitive processes and can capture subtle changes in the early stage of cognitive deficits.

This thesis investigates speech and language-based cognitive screening, addressing critical limitations in existing research, including the reliance on spontaneous speech elicited through structured tasks that introduce cognitive burden, the scarcity of longitudinal studies crucial for early intervention, the predominant use of small, balanced datasets that fail to reflect real-world diagnostic distributions and the lack of deployable tools for widespread and casual testing.

Leveraging the longitudinal biographical interviews from the ILSE dataset, this work investigates both state screening (detecting current cognitive decline) and predictive screening (forecasting cognitive decline up to 12 years in advance), which is crucial for early intervention. This work introduces a comprehensive framework for processing and analyzing longitudinal, naturalistic datasets, including methods for addressing data imbalance, exploring feature selection techniques, evaluating classifier performance, and enhancing model interpretability through feature importance analysis. These methodologies provide a foundation for future research on longitudinal and naturalistic datasets.

Another key contribution of this work is the development of DemVis, an end-to-end system that integrates automated speech processing, feature extraction, and classification models to detect AD and AACD based on natural, conversational speech. DemVis offers an accessible and scalable screening tool that supports clinicians in diagnosis while also enabling casual self-assessment of both current cognitive status and future risk of cognitive decline.

Additionally, a robust multi-processing data pipeline is developed to efficiently manage large-scale speech data for speech and language-based analysis. Its modular components can be reused or adapted for future research on longitudinal and naturalistic datasets.