Multi-Task Learning for Alzheimer's Disease Progression: Biomarker Assessment and Temporal Modelling

Multi-task learning (MTL) provides a principled paradigm for predicting individual Alzheimer's Disease (AD) progression from neuroimaging data. Its predictions can be traced to individual brain regions, enabling biomarker assessment. However, current approaches face two critical limitations: feature selection instability that undermines biomarker reproducibility in high-dimensional, correlated feature spaces; and rigid temporal modelling assumptions that cannot accommodate diverse, non-linear individual trajectories. This thesis addresses both limitations within a regularised MTL framework.

A systematic empirical analysis of regularised MTL first establishes a reliable evaluation baseline and confirms both limitations. To address feature selection instability, randomised feature learning with stability selection is introduced, separating genuinely predictive features from spurious high-dimensional correlations to yield a compact and reproducible biomarker set. To move beyond isolated biomarkers, a spatio-temporal data augmentation strategy coupled with tensor decomposition then models pairwise co-atrophy relationships among brain regions, producing representations that outperform conventional region-of-interest features. To overcome rigid temporal modelling assumptions, data-driven task relationship discovery replaces fixed temporal smoothness constraints with structures learned directly from data, enabling the model to capture flexible, potentially asymmetric inter-task dependencies.

Experiments on the Alzheimer's Disease Neuroimaging Initiative (ADNI) longitudinal cohort indicate improved prediction and more stable biomarker identification under the reported settings, while the identified biomarkers remain consistent with established AD neuroanatomical patterns.

Overall, these contributions indicate that data-driven discovery of spatial biomarker interactions and temporal task dependencies yields both more accurate predictions and more interpretable, stable biomarker identification for heterogeneous AD trajectories.