Cortical thickness estimation performed in-vivo via magnetic resonance imaging (MRI) is an effective measure of brain atrophy in preclinical individuals at high risk for Alzheimer's disease (AD). However, the high dimensionality of individual cortical thickness data coupled with small population samples make it challenging to perform cortical thickness feature selection for AD diagnosis and prognosis. Thus far, there are very few methods that can accurately predict future clinical scores using longitudinal cortical thickness measures. In this paper, we propose an unsupervised dictionary learning algorithm, termed Multi-task Sparse Screening (MSS) that produces improved results over previous methods within this problem domain. Specifically, we formulate and solve a multi-task problem using extracted top-p significant features from the Alzheimer's Disease Neuroimaging Initiative (ADNI) longitudinal data. Empirical studies on publicly available longitudinal data from ADNI dataset (N = 2797) demonstrate improved correlation coefficients and root mean square errors, when compared to other algorithms.