Alzheimer’s Disease is projected to affect over 139 million people worldwide by 2050, yet most
patients are diagnosed after significant brain changes have already occurred. Early detection
expands treatment options, supports family planning, and reduces caregiver burden, but
current diagnostic pathways remain inconsistent, expensive, and frequently late.
This thesis evaluates whether machine learning applied to multimodal clinical data can improve
early prediction across the Alzheimer’s continuum. Using baseline data from approximately
2,100 participants in the Alzheimer’s Disease Neuroimaging Initiative (ADNI), we built and
tested predictive models across three diagnostic stages: cognitively normal, mild cognitive
impairment (MCI), and dementia, incorporating dozens of features spanning cognitive
assessments, neuroimaging metrics, and fluid biomarkers.
A Random Forest classifier achieved high overall accuracy (approximately 89% in preliminary
analyses), with about one third of MCI patients in the cohort progressing to dementia,
underscoring the clinical urgency of identifying this transitional group. The strongest predictor
was the Clinical Dementia Rating Sum of Boxes, a measure of everyday functional decline,
followed by the Functional Activities Questionnaire and digital cognitive composites. These
results align with current clinical practice, suggesting that machine learning can augment rather
than replace clinician judgment.
The work also addresses real world data constraints. Initial class imbalance reduced sensitivity
for minority diagnostic groups (roughly 40–50% before rebalancing), and over threequarters of
cerebrospinal fluid biomarker entries were missing, a common reality in clinical datasets that
any deployed model must accommodate.
This session will present our methodology, key findings (with final metrics to be confirmed), and
broader implications for designing AI tools that perform reliably in clinical environments where
data is incomplete, imbalanced, and variable.
