Single-view, video-based diagnosis of Parkinson's Disease based on arm and leg joint tracking

Abstract

Automatic diagnosis of Parkinson's Disease (PD) from sensor data is an important topic given the growing numbers of patients, and the increasing costs to the quality of life of an aging society. Several approaches have been proposed aimed at such an automatic diagnosis, but often suffer from complicated sensor setups or setups ill-fitting for the limitations in clinical settings. Here, we present a system that uses frequency-based analysis of joint data from both arms and legs from a single, frontally-viewed video of people walking towards a camera. We evaluate three machine-learning models on frequency-based features extracted from the joint dynamics on two larger datasets containing a total of N=300 videos of over 50 PD patients and healthy control people. Results confirm typical clinical expectations (leg frequencies are slower in PD patients) and in addition show excellent generalizability even across datasets with performance of up to 97% for an Ensemble classifier.