Use of pattern classification to identify mild cognitive impairment and predict cognitive decline

Research Doctorate - Doctor of Philosophy (PhD)Alzheimer’s disease (AD) is the most common cause of dementia in older people, with the prevalence doubling for every 5-year interval beyond the age of 65. The combination of our ageing society and the broad-reaching and devastating impacts of AD make research into this disease an urgent priority. A desirable aim of such research is to develop means of making early and accurate diagnoses of individuals who either have or are at increased risk of developing AD. This will allow for timely interventions that may be effective in managing or treating AD. There is a cognitive continuum from normal ageing to dementia, with mild cognitive impairment (MCI) being a syndrome widely considered to be a prodromal stage of dementia. In this thesis, pattern classification algorithms were used for the identification of MCI and prediction of decline from normal cognition to MCI, and MCI to AD. Three studies were conducted: (1) identification of amnestic MCI among community-dwelling elderly adults, (2) prediction of the transition from normal cognition to MCI in community-dwelling elderly adults, and (3) prediction of the conversion from amnestic MCI to AD in a clinic-based sample. Due to there being only subtle brain changes in the very early stages of cognitive decline, early diagnosis is particularly challenging. The first study investigated the automated detection of MCI using a combination of spatial atrophy and white matter alterations, as changes in both brain structure and the capacity for information flow within and between structures are important contributors to cognitive dysfunction. Additionally, numerous socio-demographic, lifestyle, health and other factors were implicated in the misclassification of individuals. The second study used neuropsychological test scores and neuroimaging morphological measures to identify cognitively normal individuals at increased risk of developing MCI, and appears to be the first study to use pattern classification methods for this purpose. The third study investigated conversion from MCI to AD using multimodal data that included cerebrospinal fluid (CSF) protein concentrations, neuroimaging, and neuropsychological test scores. The classification and prediction schema used in these studies comprised feature extraction, feature selection and classification stages. Using an automated feature extraction process, measurements of brain structures were computed from neuroimages. In addition to these, cognitive data obtained from neuropsychological assessments and CSF biomarker data were also used. Meaningful features, which enabled optimal differentiation between cognitive groups, were then identified from the range of neuroimage, neuropsychological and CSF biomarker features using a feature selection process. In the classification stage, non-linear support vector machines were then used to train classifiers and test classification performance. These pattern classification methods achieved a high level of performance in all three studies. In addition, performance was enhanced by using a combination of multiple data modalities over any one modality alone. The use of the scheme to identify discriminating markers enhances the current understanding of AD progression. Also importantly, the scheme has the potential to detect MCI in the early stages of its development. Early detection would enable interventions designed to prevent or slow the development of AD and other dementias to begin as soon as possible