A Machine Learning Architecture to detect Alzheimer’s Disease Progression based on the Evaluation of Cognitive and Functional Attributes of Neuropsychological Assessments

With an aging population and the increased burden and economic cost of dementia on the community, coupled with an ongoing pandemic, it is becoming critical to develop a faster, cheaper, reliable way of diagnosing and screening for dementia and its progression. Dementia is a condition associated with memory decline, cognitive impairment, and difficulties in language, problem-solving, and sometimes functional impairment. Pre-diagnosis of common dementia conditions such as Alzheimer’s disease (AD) in the initial stages is crucial to help in early intervention, treatment plan design, disease management, and for providing quicker healthcare access. Current pathological assessments are often physically invasive, psychologically stressful, and their availability in poor countries and rural areas is low. In addition, many neuropsychological assessments are time-consuming, rarely cover all cognitive domains involved in AD diagnosis, and they do not measure the individual’s cognitive and functional abilities together. Therefore, the design and implementation of an intelligent method for AD progression from few cognitive and functional items in a manner that is accessible, easy, affordable, quick to perform, and does not require special and expensive resources, is desirable. This thesis investigates the issue of AD progression based on cognitive and functional items using machine learning to offer good performance (accuracy, time, etc.) and accessibility besides providing valuable knowledge for clinicians during the clinical assessment.

The thesis proposes a Machine Learning Architecture for Alzheimer’s Disease Progression (MLA-ADP), which contains a novel classification algorithm called Alzheimer’s Disease Class Rules (AD-CR). The proposed AD-CR algorithm learns models from the distinctive feature subsets that contain rules with low overlapping among their cognitive and functional items yet are easy to be interpreted by the clinicians during clinical assessment. More importantly, our research investigated cognitive elements, functional abilities and their correlations during dementia progression using computational intelligence, and it was able to identify sets of key cognitive and functional items within neuropsychological assessments. The cognitive items mainly covering the cognitive domains of learning and memory, and language, when processed by machine learning techniques, produced models that performed well and showed superior models using the AD-CR algorithm. As for functional items, demographic features must be included with functional items for AD progression detection, at least when using machine learning techniques. Overall, cognitive items appear to be more influential, the AD-CR algorithm model was still able to generate results across the dissimilar evaluation metrics that are within the standard medical research.

To measure the performance of the dementia progression models, extensive experiments have been conducted using classification methods on the Disease Neuroimaging Initiative data repository (ADNI) datasets. Results obtained by ten-fold cross validation showed that fewer cognitive and functional items can be processed by the AD-CR algorithm generating models that maintain adequate performance in terms of accuracy, sensitivity, and specificity. The models derived by the AD-CR algorithm are competitive in terms of accuracy, specificity, and sensitivity rates. For the cognitive subsets, processing three items by the AD-CR algorithm with the addition of demographic information derives models with 91.25% accuracy, 89.50% sensitivity, and 92.90% specificity. Whereas for functional items, when four items are processed by the AD-CR algorithm, the models derived show 87.57% accuracy, 86.70% sensitivity, and 88.30% specificity.