Parkinson's disease (PD) is heterogeneous: it has many subtypes and the disease progresses at different rates in different subtypes. Disease progression in early-stage PD is observed as signal changes in SPECT imaging with 123I-FP-CIT, which is called DaTscan imaging, or simply DaTscan. The goal of the research proposed here is to create accurate models of heterogeneous PD progression using DaTscan images. Such models will not only provide additional insight into PD, but they are also critically important in assessing the effect of neuro-protective therapy. A new set of models called mixtures of linear dynamical systems (MLDS) are proposed to model early-state PD progression as it manifests in DaTscans. MLDS models combine machine-learning methods with linear dynamical system theory. They capture many features of early-stage PD progression: laterality, non-linear progression, as well as PD heterogeneity. Preliminary results show that, MLDS is accurate, finds progression subtypes, relates well to clinical data (MDS-UPDRS motor scores), and gives genuinely new insights about PD progression. The proposed research aims to develop the MLDS methodology in region-of-interest as well as voxel-based frameworks. A detailed discussion of the MLDS theory, model fitting, and the relation to clinical data is included. Longitudinal DaTscan images as well as MDS-UPDRS motor scores are available for over 440 subjects from the Parkinson's Progression Markers Initiative (PPMI), and this data set will be used along with MLDS to create the PD progression models.
How Parkinson's disease (PD) progresses is poorly understood. This research proposes to develop precise disease progression models for PD using machine-learning. These models will give greater insight into PD progression and also help assess PD therapy.