Longitudinal non-negative matrix factorization identifies the altered trajectory of motor symptoms in Parkinson's disease.

Parkinson's disease (PD) is the second most common neurodegenerative disease with progressive structural alterations throughout the brain, resulting in motor symptoms that seriously affect patients' daily life. The present study then aimed to explore the progressive co-changes in gray matter patterns in PD and identify the longitudinal neuroimaging biomarkers that could predict the progressive motor symptoms of PD. Non-negative Matrix Factorization (NMF) was first used to decompose gray matter images into 7 latent factors from healthy samples, and then the latent factors were validated on an independent dataset to verify the stability of the structural factors. Parkinson's patients (including baseline, 1-year follow-up, and 2-year follow-up data) and healthy controls (HC) from Parkinson's Progression Markers Initiative (PPMI) were used to find the correlation between factor weights and motor-symptom related Movement Disorder Society Unified Parkinson's Disease Rating Scale (MDS-UPDRS) scores. The decreasing trend of the factor weights with increasing disease duration was found in the first 6 factors. The XGBoost prediction model demonstrated that Factor 2 (motor function), 3 (perceptual processing) & 7 (cerebellum) played pivotal roles in longitudinally predicting MDS-UPDRS-Ⅱ scores, whereas Factor 3 & 5 (subcortical basal ganglia) accounted for most change in MDS-UPDRS-Ⅲ. Our research indicated that the NMF factors could capture the progressive alterations of structural architectures in PD, and the factor weights were capable of predicting the clinical motor symptoms. This provides new perspectives for exploring the neural mechanisms underlying the disease and future clinical diagnostic and therapeutic approaches associated with disease progression.