Parkinson’s disease (PD) is a neurodegenerative disorder being the characterization of motor gait impairments, dysarthria (speech disorders) and hypomimia (facial rigidity), key for access a reliable diagnosis and treatment planning. Nevertheless, the characterization of such symptons is based on expert observations and pose significant challenges because remarked variability and phenotyping of the disease. Furthermore, the analysis of independent symptoms impact in delayed diagnosis and treatments mismanagement. This study proposes a non-invasive, and multimodal strategy that combines gait and facial expression to support PD classification. Firstly, anatomical reference points are computed from gait and facial expressions to coding compact and discriminative Riemannian covariance descriptors. Gait features are modeled using quaternions, while facial expressions are described through kinematic features. Then, independent Riemannian branchs were learned for each modality, which were fused in a late Euclidean layer to recover and unified PD classification. The model was tested on a dataset of 580 video recordings from 11 PD patients and 18 control subjects, achieving an accuracy of 92% ± 0.02, and AUC of 90% ± 0.03. These results demonstrate that integrating gait and facial expression features, following a Riemannian-based neural classification, offers significant potential for assessment tools in clinical settings.