Vision-Based Assessment of Parkinsonism and Levodopa-Induced Dyskinesia
Keywords: Parkinson disease, vision-based, Levodopa-Induced Dyskinesia, motor, Parkinson disease assessment.
Overview of Research
Parkinson’s disease (PD) is the second most common neurodegenerative disorder, affecting more than 4 million people worldwide. Persons with PD experience a host of motor symptoms including tremor, bradykinesia (slowness of movement), rigidity, and freezing of gait. Since its discovery in the 1970s, levodopa has been the gold standard treatment for PD symptoms. However, its usefulness is hampered by the development of motor complications after prolonged usage. These motor complications are called levodopa-induced dyskinesias (LID), and present as involuntary, irregular motions that flow from one body part to another or contortion of limbs into abnormal positions. As a result, selection of drug regimens is highly personalized and often composed of multiple drugs, seeking to maximize antiparkinsonian benefits while minimizing dyskinesia. Patients attend regular clinic visits, where clinicians use rating scales to record characteristics of PD motor signs (e.g. anatomical distribution, duration, functional impact). However, these rating scales are inherently subjective and can be significantly influenced by rater experience. Furthermore, the intermittent nature of clinic visits can fail to capture important changes in a patient’s condition.
The proposed solution is to use computer vision for objective assessment of parkinsonism and LID. Using state-of-the-art pose estimation algorithms based on deep learning, body movements were extracted from videos of Parkinson’s clinical assessments. Movement features (e.g. speed, acceleration, periodicity) were computed from movement trajectories and used to predict the severity of involuntary movements on clinical rating scales. The contributions of this study are as follows:
- Evaluation of pre-trained deep learning algorithms for human pose estimation in videos of clinical assessments
- Comprehensive exploration of movement features extracted from video analysis and their correlation to clinical ratings of PD and LID severity
- Development and testing of a system for detecting the presence of PD/LID and their respective severities, as well as identification of the most important movement features for good performance
- Exploration of the clinical utility of objective features as a new scoring paradigm for PD assessment
This study provides the first detailed analysis of the performance of off-the-shelf deep learning algorithms for vision-based assessment of PD, and indicates promising potential for computer vision and deep learning to be applied to PD clinical practices. The long term plan is the development of a mobile/tablet application for automated PD assessment to support clinicians and empower patients to actively contribute to management of their symptoms.
An anonymized dataset containing movement trajectories and clinical scores has been made available at: https://github.com/limi44/Parkinson-s-Pose-Estimation-Dataset
Publications
M. H. Li, T. A. Mestre, S. H. Fox, B. Taati, “Vision-Based Assessment of Parkinsonism and Levodopa-Induced Dyskinesia with Deep Learning Pose Estimation,” arXiv:1707.09416 [cs], July 2017. Submitted to IEEE Journal of Biomedical and Health Informatics on August 1st, 2017.
M. H. Li, T. A. Mestre, S. H. Fox, B. Taati, “Automated Vision-Based Analysis of Levodopa-Induced Dyskinesia with Deep Learning,” presented at the 2017 37th International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC).
M. H. Li, “Objective Vision-Based Assessment of Parkinsonism and Levodopa-Induced Dyskinesia in Persons with Parkinson’s Disease,” M.ASc Thesis, University of Toronto, Toronto, Canada, 2017.
Research Team
Michael Li, University of Toronto
Tiago Mestre, University of Ottawa
Susan Fox, University of Toronto
Babak Taati, Toronto Rehabilitation Institute