Supervised Machine Learning Assessment of Dementia Using Feature Selection Filter Methods

Abstract

The prevalence of dementia is increasing globally. Due to the massive resources required, this issue is pressuring governments and private healthcare systems. Accurate diagnosis by clinicians on the cause of dementia, such as Alzheimer’s disease (AD), is difficult because of the time and assessments needed like neuropathological. The issue becomes more challenging when considering if various brain lesions contribute to the pathological assessment of dementia, the relationship of these lesions to the various dementia conditions, how they interact, and how to quantify them. Thereby, systematically assessing neuropathological measures by their degree of association with dementia, especially AD, may lead to better diagnostic systems and treatment targets. One promising approach that can answer these challenges is to develop data-driven solutions with core functions of feature evaluation and automatic subject classification based on machine learning (ML). Recent research studies in medical diagnosis, including dementia research, reveal that ML techniques, when used with feature selection, can identify critical features of Alzheimer-related pathologies and their association with the disease’s diagnosis and prognosis. The feature selection removes noisy features from the dementia data to increase the predictive performance and improve interpretability while reducing the dimensionality and computational complexity. However, filter-based feature selection methods can generate dissimilar feature rankings and may be sensitive to the correlations among themselves. This thesis investigates dementia with a focus on AD neuropathological assessments from a data-driven perspective to develop mechanisms to assist pathologists during these clinical assessments. The thesis investigation comprises phases such as feature ranking, feature-feature correlation, and classification. The work determines the impact of neuropathological feature-features correlations on the feature ranking for better biomarker identification. The investigation assesses real datasets related to dementia, the Cognitive Function and Aging Studies (CFAS) and the Alzheimer’s Disease Neuroimaging Initiative (ADNI), using filter methods and classification techniques. The results showed that classification models generated from the CFAS and ADNI sets of chosen neuropathological features were strong in terms of sensitivity, accuracy, and other measures when mined by different classification techniques. In the ADNI dataset results, the significant neuropathological features contributing to AD included neocortical neuritic plaques, Braak stage, Thal phase, diffuse plaques, and cerebral amyloid angiopathy (CAA), all of which showed a high correlation with AD’s diagnostic label. In the CFAS dataset, the results were consistent with those derived from the ADNI dataset. Moreover, among the filter methods considered, reliefF had the strongest correlation with feature-feature correlations in both ADNI and CFAS datasets, less sensitive to feature-feature correlations. However, no filter method had clear dominance over ADNI results. More essentially, the results indicated limited consistency in feature rankings between ADNI and CFAS. However, reliefF had the most agreement, while the Gain Ratio method had less consistency in ranking the features in both datasets. In summary, this thesis provided valuable insights into the application of filter methods and neuropathology data for developing classification models for dementia conditions’ diagnosis. The study demonstrated the significance of considering feature-feature correlations when selecting influential features and the impact of different filter methods on feature ranking and classification performance. These findings suggest that the proposed approach could effectively minimise the discrepancy of feature ranking and generate an impactful set of features for classification algorithms. These results had practical implications for pathologists in improving the understanding of AD pathology. Furthermore, the study has highlighted the potential for future research to leverage diverse filter methods to identify more reliable biomarkers and enhance the detection of dementia, particularly for AD.