| dc.description.abstract | Alzheimer's disease (AD) is a complex neurodegenerative disorder characterized by memory loss, together with cognitive deficits affecting language, emotional affect, and interactional communication. Diagnosis and assessment of AD is formally based on the judgment of clinicians, commonly using semi-structured interviews in a clinical setting. Manual diagnosis is therefore slow, resource-heavy, and hard to access, so many people don't get diagnosed - and therefore using some kind of automatic method would help. Using the most recent advances in deep learning, machine learning, and natural language processing, this thesis empirically explores how content-free, interaction patterns are helpful in developing models capable of identifying AD from natural conversations with a focus on particular phenomena found useful in conversational analysis studies. The models presented in this thesis use lexical, disfluency, interactional, acoustic, and pause information to learn the symptoms of Alzheimer's disease from text and audio modalities. This thesis comprises two parts. In the first part, by studying a conversational corpus, we find there are certain phenomena that are really strongly indicative of differences between AD and Non-AD. This analysis shows that interaction patterns are different between an AD patient and a Non-AD patient, including types of questions asked from patients, their responses, delay in responses in the form of pauses, clarification questions, signaling non-understanding, and repetition of questions. Although it is a challenging problem due to the fact that these dialogue acts are so rare, we show that it is possible to develop models that can automatically detect these classes. The second part then shifts to look at AD diagnosis itself by looking into interactional features including pause information, disfluencies within patients speech, communication breakdowns at speaker changes in certain situations, Ngram dialogue act sequences. We found out that there are longer pauses within the AD patients utterances and more attributable silences in response to questions as compared to Non-AD patients. It also showed that using different fusion techniques with speech and text modality has maximise the combination and use of different feature sets showing that these features/techniques can give quite good accurate and effective AD diagnosis. These interaction patterns may serve as an index of internal cognitive processes that help in differentiating AD patients and Non-AD patients and may be used as an integral part of language assessment in clinical settings. | en_US |
