MV-DEN: A Geometry-Driven Latent Manifold Framework for Continuous Modelling of the Prodromal-to-Parkinson's Disease Continuum

The challenge of modelling the progression from the prodromal phase to clinically defined Parkinson's disease (PD) is a major challenge of the existing approaches that either (i) adopt a definition of discrete disease stages, (ii) predict clinical scores by black-box regression approaches. Clearly, there is no unified framework that jointly describes the prodromal-to-PD continuum as a continuous multivariate disease. Accordingly, we propose the concept of a Multivariate Disease Embedding Network (MV-DEN), which comprises a geometry-inspired approach to modelling the continuum of disease progression from the prodromal phase to the clinical state of PD using longitudinal structural MRI. To achieve this, the study utilised a balanced dataset comprising 520 subjects (260 PD and 260 prodromal) to learn a two-dimensional latent disease manifold from subcortical volumetric features and MDS-UPDRS and MoCA scores. PD cases within the learned latent space in order to quantify the severity levels (i) continuously, (ii) relatively to the population, (iii) geometrically over time, and (iv) patient-specific velocities. Therefore, the proposed model demonstrated high discrimination potential, as evidenced by validation results: AUC = 0.9678, accuracy = 0.9080, and F1 score = 0.8889.