Adaptive Prior Correction in Alzheimer’s Disease Spatio-Temporal Modeling via Multi-task Learning

Multi-task learning methods have been studied in Alzheimer’s disease for cognitive status prediction and neuroimaging feature identification widely by utilizing prior constraints. However, the existing models do not explicitly model the spatio-temporal connectivity for the lack of samples and prior medical knowledge. In this article, we propose a sparse multi-task learning model for cognitive status prediction, which is adaptively weighted in sparse prior to prevent the error in spatial feature correlation learning, and incorporated with prior domain knowledge to estimate the progression with adaptive correction. Inference in our spatio-temporal model is based on majorization-minimization optimization guaranteed convergence properties. The proposed model is applied to a real-world neuroimaging study to predict cognitive tests scores and structured feature mining with MRI scans. The effectiveness of the proposed progression model is demonstrated by its superior prediction performance over multiple competing methods and accurate identification of compact sets of cognition-relevant biomarkers.