Detecting patterns of coordinated brain change over time via orthogonally projective non-negative matrix factorization and structural covariance analysis / Ανίχνευση προτύπων συντονισμένης εγκεφαλικης αλλαγής στο πέρασμα του χρόνου μέσω ορθογωνικά προβολικής μη-αρνητικής παραγοντοποίησης πίνακα και ανάλυση ανατομικής συνδιακύμανσης

Abstract

Large-scale longitudinal brain imaging studies provide unprecedented opportunities for understanding disease processes and developing early diagnostic and predictive biomarkers. However, commonly used techniques, such as voxel-based analyses, are frequently mass univariate, and fail to capture patterns of subtle, yet coordinated brain change over time, which might result from an underlying neuropathologic process. Orthogonally projective non-negative matrix factorization (opNMF) has previously shown great potential as a data-driven, interpretable dimensionality reduction and parts-based decomposition method. Its use in longitudinal studies has not yet extensively explored. Importantly, direct application of NMF methods to images that contain both positive and negative parts, is not straightforward. Brain changes can be such signals (e.g. simultaneous decrease of gray or white matter regional volumes and increase in white matter hyperintensities of CSF). Here, we propose a semi variant of opNMF (semi-opNMF) and a modification of the input for the original opNMF that fill in this methodological gap, in the context of mapping longitudinal brain change. Both the semi-opNMF model and the modified standard opNMF learn and extract parts-based representation that is driven by age-related longitudinal patterns of structural covariance (LPSCs) in data, under the hypothesis that such patterns might reflect evolving neuropathological processes that affect brain regions simultaneously. We have empirically proven that the semi-opNMF model can quickly converge to a global or local optimum and obtain high sparsity for clinical interpretability. The extracted LPSCs were highly correlated with various clinical measures in ADNI and BLSA. Additionally, LPSCs had comparable predictive power with some of the state-of-the-art biomarkers of MCI and AD available and they had higher power in classifying progressive MCI and static MCI patients than other widely used extracted regions. The proposed model shows great potential in longitudinal analysis with mixed-sign signals and promotes clinical interpretability.