Empirical analysis of regularized multi-task learning for modelling Alzheimer’s disease progression

Recently, there have been a wide spectrum of ma-chine learning models developed to model Alzheimer’s disease (AD) progression. Multi-Task Learning (MTL) approaches has been commonly used by these studies to address challenges of missing and insufficient AD data. Typical MTL studies in AD focuses on obtaining high quality of baselines (MRI features and cognitive scores) from AD raw data and exploring advanced re-gression models for exploring their relationship and correlations. These studies follow a unified regularized MTL framework to process AD datasets with simple evaluation matrix. But another easy-ignorable issue here is whether experimental evaluation strategies are objective and reliable to access MTL performance. There is little attention on studying how to design feasible experimental protocols and evaluation matrix for assessment of regularized MTL models. In this paper, we describe an empirical study and analysis that investigate above question. Four typical structural regularization approaches in MTL study are examined, including (Ridge, Lass, TGL and cFSGL) [1],[2]. Four issues affecting evaluation process of regularised MTL models are evaluated by experiments: 1) evaluation indicators, 2) repeated experimental times; 3) size and portion of training data; 4) number of tasks in MTL. The results demonstrate that regularized MTL models like cFSGL are capable of predicting AD progression with high accuracy, in many challenging cases of data missing, insufficiency or even single MRI data input. One important finding is the performance gain of cFSGL may not only from its ability on dealing with sparsity of AD feature data labels. It is more likely due to existence of a low rank space inside original AD data features. We also discover and proof some limitations of regularized MTL in AD study: the assumption of temporal smoothness in regularized MTL models for AD study limits their performance improvement of the initial task. It is a special relationship that fails to accurately capture certain tasks. Some MTL models like cFSGL have great potential of improvement at late stage prediction of AD progression.