PD motor symptoms arise when patients reach 60% or more dopaminergic loss, and this prodromal phase can take up to 10 years. Presently, the prodromal phase is not detectable by the current standard of care, which affects the possibility of early neuroprotective therapies use. 

To compare two machine learning models based on longitudinal structural connectomes and subcortical region volumes to properly classify three cohorts: healthy controls, Parkinson’s disease, and Prodromal subjects. 

In this study, we used a longitudinal connectome methodology based on the structural Parkinson’s-relevant connectomes on diffusion tensor imaging (DTI). Two datasets were used: the Parkinson’s Progression Markers Initiative and our clinical trial dataset. Longitudinal connectome features were used as the inputs to the machine learning algorithm to discriminate progression between those with (n = 46) and without (n = 40) Parkinson’s disease. The progression model learned was then applied to an independent group of subjects who are at risk of developing Parkinson’s (n = 34). Further, we compared our method against a model trained on volumetric changes in the brain’s subcortical regions relevant to PD.

The experiments indicated that the model identified a Parkinson’s specific neurodegenerative progression in the prodromal group using images from different centers and scanners (AUC=0.73). Also, we found that a model based on longitudinal connectomes outperformed a model based on the volumetric changes derived from MRI images (p=0.03), thereby highlighting the advantage of structural connectivity to track disease progression.

Tools to identify early onset of Parkinson’s disease are an urgent medical need for diagnosis and treatment. In this study, we tested the robustness of our longitudinal connectome method on two separate datasets acquired by multiple centers and scanners. By understanding identifying subjects who are at risk to develop a neurodegeneration process, we enable these patients to be treated earlier.