Developing neuroimaging biomarkers for neurodegeneration and ageing using machine learning methods

In this thesis, we present three novel methods based on machine learning for use with MRI-derived neuroimaging data, all with the aim of aiding biomarker development for neurodegeneration. Resting-state functional MRI (rfMRI) can potentially detect early functional changes in disease. Therefore, the first method is a novel supervised learning algorithm for use in classifying rfMRI data into two groups. The main advantage over existing rfMRI-based classification approaches is that the entire voxel by time data is fed in without any prior decomposition or parcellation of the data into brain regions, and it does not require any prior knowledge of potential discriminatory networks. We show that the algorithm can give interpretable results for simulated data, performs better than two existing approaches for a Parkinson's disease dataset and gives results consistent with those reported previously for an anaesthesia dataset. Ageing is strongly linked with neurodegeneration. Hence, the second method is an errors-in-variables model for estimating the brain age of individuals. It takes a fundamentally different approach compared to existing methods, which gives it the advantages of being able to capture inverted U-shaped brain ageing trajectories, provide visualisable trajectories and potentially model interactions with disease. Using simulated data, we show that it is effective at estimating brain age, and can match and in some cases outperform current approaches. Using the UK biobank dataset, we show that the difference between chronological age and predicted brain age for an individual correlates with variables that are consistent with those reported previously. We also show that it can potentially provide information that is complementary to existing methods. The final method is a tool for decomposing multimodal neuroimaging data into spatial maps and associated trajectories. The main differences compared to existing methods are that the decomposition is driven by a variable of interest, which in our case was age, and each brain region is associated with only one trajectory as opposed to a mixture of different trajectories. This means that the components do not have to be associated with age post-hoc and the results are potentially more interpretable. Using simulated data, we show that, compared to two existing methods, our method can more accurately reproduce trajectories and associated spatial maps. Using the UK biobank dataset, we show that our method estimates trajectories and associated regions that are in agreement with previously published studies.