What the trained eye cannot see: Quantitative kinematics and machine learning detect movement deficits in early-stage Parkinson's disease from videos

Abstract

Background

Evaluation of disease severity in Parkinson's disease (PD) relies on motor symptoms quantification. However, during early-stage PD, these symptoms are subtle and difficult to quantify by experts, which might result in delayed diagnosis and suboptimal disease management.

Objective

To evaluate the use of videos and machine learning (ML) for automatic quantification of motor symptoms in early-stage PD.

Methods

We analyzed videos of three movement tasks—Finger Tapping, Hand Movement, and Leg Agility— from 26 aged-matched healthy controls and 31 early-stage PD patients. Utilizing ML algorithms for pose estimation we extracted kinematic features from these videos and trained three classification models based on left and right-side movements, and right/left symmetry. The models were trained to differentiate healthy controls from early-stage PD from videos.

Results

Combining left side, right side, and symmetry features resulted in a PD detection accuracy of 79 % from Finger Tap videos, 75 % from Hand Movement videos, 79 % from Leg Agility videos, and 86 % when combining the three tasks using a soft voting approach. In contrast, the classification accuracy varied between 40 % and 72 % when the movement side or symmetry were not considered.

Conclusions

Our methodology effectively differentiated between early-stage PD and healthy controls using videos of standardized motor tasks by integrating kinematic analyses of left-side, right-side, and bilateral symmetry movements. These results demonstrate that ML can detect movement deficits in early-stage PD from videos. This technology is easy-to-use, highly scalable, and has the potential to improve the management and quantification of motor symptoms in early-stage PD.