Video-based automatic Early Parkinson's disease detection system using biomechanical features

Changes in biomechanical features are an important early symptom of Parkinson's disease(PD). Intending to make Parkinson's disease diagnosis easier and more accessible, this paper presents a marker-less system that performs automatic detection and classification of early Parkinson's disease, based on the evaluation and inference of biomechanical features using a single 2D video camera and naive Bayesian algorithm.9 biomechanical features are selected according to the statistic analysis result of Parkinson's disease and Unified Parkinson Disease Rating Scale (UPDRS), including two types of features, one is spatial biomechanical features(step length, range of joint motion, etc.) another is temporal biomechanical features(cadence and gait cycle phase). Movement information of individual joints is extracted from videos of subjects' sagittal plane using Openpose, which is a pose estimation algorithm built with deep learning. Parameters of the Bayesian classification model are estimated by symptom statistics of Parkinson's disease to avoid a lack of generalization of the classification model due to small sample sizes. Experiment results show that the achieved accuracy of the probabilistic classification was 95.7%.