Machine learning-based approach to help diagnosing Alzheimers disease through spontaneous speech analysis

Abstract

Alzheimer’s disease and other dementias have been recognized as a major public health problem among the elderly in developing countries. We address this issue by exploring automatic noninvasive techniques for diagnosing patients through analysis of spontaneous, conversational speech. The technique we are proposing is a variant of n-gram based kNN machine learning technique. Since we use byte-level n-grams, we do not use any language dependent information, including word boundaries, character case, white-space characters or punctuation. Twelve adults diagnosed with dementia of Alzheimer type (DAT) participate in the study. All DAT participants were interviewed at adult day care center for people with Alzheimer’s disease or dementia in Novi Sad, the only institution of its kind in Serbia. All interviews were audio-recorded, transcribed verbatim by a trained researcher, and checked for accuracy by the authors. Means for the Mini-Mental Status Exam distinguished the two groups: moderate and mild. Our plan is to compile a control dataset based on the interviews of healthy elderly that do not differ significantly in age, sex or education level from the DAT participants. We plan to compare DAT and healthy elderly participants to test how well our techniques will discriminate between these groups. In this paper, we make some preliminary distinction between the two groups of the DAT participants. Our plan is to develop new, more sophisticated classification techniques, based on Machine Learning and Natural Language Processing. We hope that our techniques will show promising as diagnostic and prognostic additional tools that may help earlier diagnosis of DAT and determining its degree of severity.