Investigation of Multi-dimensional Tensor Multi-task Learning for Modeling Alzheimer's Disease Progression

Machine learning (ML) techniques for predicting Alzheimer's disease (AD) progression can significantly assist clinicians and researchers in constructing effective AD prevention and treatment strategies. The main constraints on the performance of current ML approaches are prediction accuracy and stability problems in medical small dataset scenarios, monotonic data formats (loss of multi-dimensional knowledge of the data and loss of correlation knowledge between biomarkers) and biomarker interpretability limitations. This thesis investigates how multi-dimensional information and knowledge from biomarker data integrated with multi-task learning approaches to predict AD progression. Firstly, a novel similarity-based quantification approach is proposed with two components: multi-dimensional knowledge vector construction and amalgamated magnitude-direction quantification of brain structural variation, which considers both the magnitude and directional correlations of structural variation between brain biomarkers and encodes the quantified data as a third-order tensor to address the problem of monotonic data form. Secondly, multi-task learning regression algorithms with the ability to integrate multi-dimensional tensor data and mine MRI data for spatio-temporal structural variation information and knowledge were designed and constructed to improve the accuracy, stability and interpretability of AD progression prediction in medical small dataset scenarios. The algorithm consists of three components: supervised symmetric tensor decomposition for extracting biomarker latent factors, tensor multi-task learning regression and algorithmic regularisation terms. The proposed algorithm aims to extract a set of first-order latent factors from the raw data, each represented by its first biomarker, second biomarker and patient sample dimensions, to elucidate potential factors affecting the variability of the data in an interpretable manner and can be utilised as predictor variables for training the prediction model that regards the prediction of each patient as a task, with each task sharing a set of biomarker latent factors obtained from tensor decomposition. Knowledge sharing between tasks improves the generalisation ability of the model and addresses the problem of sparse medical data. The experimental results demonstrate that the proposed approach achieves superior accuracy and stability in predicting various cognitive scores of AD progression compared to single-task learning, benchmarks and state-of-the-art multi-task regression methods. The proposed approach identifies brain structural variations in patients and the important brain biomarker correlations revealed by the experiments can be utilised as potential indicators for AD early identification.