Joint and Long Short-Term Memory Regression of Clinical Scores for Alzheimer's Disease Using Longitudinal Data

Alzheimer's disease (AD), the most common type of the dementia, is a progressive neurodegenerative disease that mainly affects elderly. It causes a high financial burden for patients and their families. For effective treatment of AD, it is important to identify the AD progression of clinical disease over time. As the cognitive scores can effectively indicate the disease status, the prediction of the scores using the longitudinal magnetic resonance imaging (MRI) data is highly desirable. In this paper, we propose a joint learning and clinical scores prediction method for AD diagnosis via longitudinal MRI data. Specifically, we devise a novel feature selection method that consists of a temporally constrained group LASSO model and the correntropy. The baseline MRI data is used to jointly select the most discriminative features. Then, we use the stacked long short-term memory (SLSTM) to effectively capture useful information in the input sequence to predict the clinical scores of future time points. Extensive experiments on the Alzheimer's disease Neuroimaging Initiative (ADNI) database are conducted to demonstrate the effectiveness of the proposed model. Our model can accurately describe the relationship between MRI data and scores, and thus it can be effective in predicting longitudinal scores.