Progression to dementia due to Alzheimer’s Disease (AD) is a long and protracted process that involves multiple pathways of disease pathophysiology. Predicting these dynamic changes has major implications for timely and effective clinical management in AD. There are two reasons why at present we lack appropriate tools to make such predictions. First, a key feature of AD is the interactive nature of the relationships between biomarkers, such as accumulation of β-amyloid -a peptide that builds plaques between nerve cells-, tau -a protein found in the axons of nerve cells- and widespread neurodegeneration. Current models fail to capture these relationships because they are unable to successfully reduce the high dimensionality of biomarkers while exploiting informative multivariate relationships. Second, current models focus on simply predicting in a binary manner whether an individual will develop dementia due to AD or not, without informing clinicians about their predicted disease trajectory. This can result in administering inefficient treatment plans and hindering appropriate stratification for clinical trials. In this thesis, we overcome these challenges by using applied machine learning to build predictive models of patient disease trajectories in the earliest stages of AD. Specifically, to exploit the multi-dimensionality of biomarker data, we used a novel feature generation methodology Partial Least Squares regression with recursive feature elimination (PLSr-RFE). This method applies a hybrid-feature selection and feature construction method that captures co-morbidities in cognition and pathophysiology, resulting in an index of Alzheimer’s disease atrophy from structural MRI. We validated our choice of biomarker and the efficacy of our methodology by showing that the learnt pattern of grey matter atrophy is highly predictive of tau accumulation in an independent sample. Next, to go beyond predicting binary outcomes to deriving individualised prognostic scores of cognitive decline due to AD, we used a novel trajectory modelling approach (Generalised Metric Learning Vector Quantization – Scalar projection) that mines multimodal data from large AD research cohorts. Using this approach, we derive individualised prognostic scores of cognitive decline due to AD, revealing interactive cognitive, and biological factors that improve prediction accuracy. Next, we extended our machine learning framework to classify and stage early AD individuals based on future pathological tau accumulation. Our results show that the characteristic spreading pattern of tau in early AD can be predicted by baseline biomarkers, particularly when stratifying groups using multimodal data. Further, we showed that our prognostic index predicts individualised rates of future tau accumulation with high accuracy and regional specificity in an independent sample of cognitively unimpaired individuals. Overall, our work used machine learning to combine continuous information from AD biomarkers predicting pathophysiological changes at different stages in the AD cascade. The approaches presented in this thesis provide an excellent framework to support personalised clinical interventions and guide effective drug discovery trials.