Automatic Detection of Early Signs of Alzheimer’s Disease in Speech and Language

Abstract

The relevance of Alzheimer’s disease (AD) is growing due to ageing population, and an increasing amount of research is being conducted in both treatment development and early detection of the disease. Previous research has shown that changes in language use could be one of the earliest signs of cognitive decline in AD, and these changes could be automatically detected using natural language processing (NLP) and artificial intelligence (AI). While NLP- and AI-based tools could contribute to detecting AD early in a non-invasive, fast, cheap, and accessible way, there is still a major gap between the scientific knowledge and its applicability to clinical practice. In the current thesis, I first conducted a systematic literature review of the studies looking at automatic speech-based AD detection and identified the key challenges in the state-of-the-art: (1) the lack of longitudinal language data; (2) the lack of replicability, generalisability, and standardisation; and (3) the lack of ethical guidelines. To tackle these issues, I first present a novel corpus of longitudinal transcripts of interview recordings with public figures, and demonstrate the usefulness of this kind of data in understanding longitudinal language changes in AD. Second, I replicate a previous case study on a larger group of individuals, explore the generalisability of the language change, and identify the most informative language features that change consistently across individual speakers. Third, I focus on the standardisation of data collection methods and investigate the role of sample length in analysing AD-related language change. Fourth, I outline the ethical considerations in AI- and NLP-based AD detection from speech and language and provide a list of suggestions that could be incorporated in the development of ethical guidelines and best practices. This work aims to address some of the main challenges in automatic speech-based AD detection and fill the gaps in the existing literature to contribute to developing robust, fair, and ethical methods for the automatic detection of early signs of AD in speech and language.