Abstract
Longitudinal designs are widely used in medical studies as a means of observing within-subject changes over time in groups of subjects, thereby aiming to improve sensitivity for detecting disease effects. Paralleling an increased use of such studies in neuroimaging has been the adoption of pattern recognition algorithms for making individualized predictions of disease. However, at present few pattern recognition methods exist to make full use of neuroimaging data that have been collected longitudinally, with most methods relying instead on cross-sectional style analysis. This article presents a principal component analysis-based feature construction method that uses longitudinal high-dimensional data to improve predictive performance of pattern recognition algorithms. The method can be applied to data from a wide range of longitudinal study designs and permits an arbitrary number of time-points per subject. We apply the method to two longitudinal datasets, one containing subjects with mild cognitive impairment along with healthy controls, the other with early dementia subjects and healthy controls. Across both datasets, we show improvements in predictive accuracy relative to cross-sectional classifiers for discriminating disease subjects from healthy controls on the basis of whole-brain structural magnetic resonance image-based voxels. In addition, we can transfer longitudinal information from one set of subjects to make disease predictions in another set of subjects. The proposed method is simple and, as a feature construction method, flexible with respect to the choice of classifier and image registration algorithm. Hum Brain Mapp 37:4385-4404, 2016. © 2016 Wiley Periodicals, Inc.