Early detection of Alzheimer's disease using machine learning and modelling: development of a novel feature selection Algorithm and ratio-based sMRI biomarker : A thesis submitted in partial fulfilment of the requirements for the Degree of Doctor of Philosophy at Lincoln University

Abstract

Alzheimer’s disease (AD) is characterized by progressive neurodegeneration, with structural brain changes often preceding measurable cognitive deficits. Early detection, particularly during mild cognitive impairment (MCI) stages, is crucial for timely intervention and disease management. Despite advances in neuroimaging and computational analytics, critical gaps remain in distinguishing disease-specific structural alterations from normal aging, identifying the most vulnerable brain regions, and developing interpretable, clinically actionable biomarkers. Conventional volumetric measures frequently fail to capture subtle, region-specific atrophy and ventricular enlargement, while high-dimensional neuroimaging data pose challenges for robust feature selection and integrative modelling. This study addresses these gaps by developing a comprehensive computational framework for early AD detection that integrates advanced feature selection, novel biomarker design, and predictive modelling. The central aim was to identify structurally vulnerable brain regions, formulate sensitive volumetric biomarkers, and enable comprehensive evaluation of disease progression. At the core of this framework is Prominent Feature Selection (PFS), a novel unsupervised, model-agnostic algorithm that preserves intrinsic data structure while effectively filtering noise and irrelevant features. Application of PFS to sMRI data consistently identified key neuroanatomical markers including the lateral-ventricle, inferior-lateral-ventricle, and 3rd-ventricles, hippocampus, and amygdala aligning with established AD pathology and outperforming conventional selection methods in both stability and interpretability. Building on these findings, a ratio-based biomarker, the Ventricle-to-Tissue Ratio (VTR) and its estimated form (VTRE), was developed to integrate ventricular enlargement and brain tissue atrophy into a single, anatomically grounded metric. VTRE inherently normalizes for inter-individual anatomical variability, enhancing sensitivity to subtle early-stage changes and providing superior discrimination between early and late MCI groups. Explainable artificial intelligence (XAI) analyses further confirmed VTRE as the most influential predictive feature, supporting its clinical interpretability. The temporal evolution of VTRE was further examined using deterministic and stochastic models. Deterministic exponential decay models captured stage and sex-specific trends, while stochastic extensions accounted for variability in individual trajectories. This integrated modelling approach validates VTRE not only as a static volumetric measure but also as a dynamic indicator of disease progression. Key findings demonstrate that PFS consistently identifies biologically meaningful regions, while VTRE effectively captures the interplay between tissue loss and ventricular expansion, surpassing conventional volumetric measures. Incorporating deterministic and stochastic modelling revealed additional insights: the acceleration of decline with disease stage, the widening of sex-related differences, and the impact of stochastic variability on individual trajectories. Composite features integrating multiple vulnerable regions further enhance discrimination between disease stages. Collectively, these contributions provide a robust, interpretable, and clinically relevant framework for early AD detection, risk stratification, and progression monitoring, establishing a pathway for scalable, data-driven translational applications in neuroimaging-based diagnostics and precision medicine In conclusion, this work demonstrates that integrating advanced unsupervised feature selection, ratio-based biomarkers, and deterministic-stochastic modelling offers a statistically rigorous, biologically grounded, and clinically interpretable approach to early AD detection. The framework addresses existing knowledge gaps, enhances mechanistic understanding of neurodegeneration, and establishes a pathway for scalable, data-driven translational applications in neuroimaging-based diagnostics.