Diagnosis of Parkinson’s Disease in Genetic Cohort Patients via Stage-Wise Hierarchical Deep Polynomial Ensemble Learning

As a neurodegenerative disease, Parkinson’s disease (PD) has gradually become common in the elderly. Effective disease diagnosis has become increasingly important, especially in the patients with mutation of PD related gene. Due to the slight changes in the brain, it is very difficult to diagnose PD by neuroimaging techniques. In order to be more effective in assisting diagnosis, we further improve the deep polynomial network (DPN) as the hierarchical stacked DPN (HSDPN) and propose a stage-wise hierarchical deep polynomial ensemble learning (SHDPEL) framework for encoding multiple features to obtain high-level feature representations of different neuroimaging segmentation in PD diagnosis. Specifically, we train different segmentation features separately in the first stage. In next stage, different combinations of feature pairs will be used to learn the correlative information between different segmentations. We further integrate all branches by using a voting ensemble strategy for the classification. A series of experiments are performed on all the neuroimaging data to demonstrate the effectiveness of this method on the publicly available Parkinson’s Progression Marker Initiative (PPMI) dataset. The experimental results show that the method can achieve remarkable results and is superior to related methods.