Abstract
Our study can be meaningful for remote assessments, especially when people have to maintain social distance, e.g., in situations such as the coronavirus disease-2019 (COVID-19) pandemic.
Introduction
The gold standard for evaluating Parkinson’s Disease (PD) motor symptoms is the MDS-UPDRS Part III scale. While many symptoms, such as tremors or speech, are easily observable via camera, two critical components—rigidity (muscle stiffness) and postural stability (balance)—traditionally require a clinician to physically touch or pull the patient.
In the context of telemedicine, these "touch-required" metrics are often skipped, which can invalidate total scores and hinder remote clinical trials. This research, published in Frontiers in Aging Neuroscience, presents a novel approach to digitizing these assessments by using standard RGB video to indirectly estimate these symptoms through touchless motions.
Methodology
The authors developed a pipeline combining computer vision and machine learning to bridge the gap between visible motion and physical rigidity.
Pose Estimation and Feature Extraction
The system utilizes HRNet (High-Resolution Net) for joint estimation to improve accuracy when limbs are straightened, a known failure point for standard pose estimators. From the resulting data, three categories of features are extracted: kinematic (speed, rhythm), position (coordinate-based signals), and angle (range of motion).
The "Indirect" Model
Instead of measuring rigidity directly, the researchers trained a LightGBM (Light Gradient Boosting Machine) to find correlations between ground-truth clinical scores and features extracted from 11 touchless motions, such as gait, toe tapping, and arising from a chair.
Key Insights and Results
The study involved 104 patients with PD, with ground-truth scores provided by on-site specialists. The models demonstrated substantial performance across the board:
- High Accuracy: All four models (Neck, Upper, Lower, and Postural Stability) achieved an absolute accuracy of ≥ 70%. When allowing for a 1-point margin of error on the 0-4 scale, accuracy soared to between 87% and 97%.
- The Proxy Power of Gait: Features from walking contributed significantly (over 13%) to all models, confirming that gait is a systemic biomarker reflecting multiple facets of motor degradation.
- Geometry over Kinematics: Surprisingly, position and angle features were more influential than kinematic features like speed. This suggests that the spatial geometry and range of a movement are better indicators of underlying muscle stiffness than how fast the patient moves.
- Efficiency: Through item minimization, the researchers found that neck rigidity could be estimated using just four motions (Finger tapping, Pronation/Supination, Gait, and Toe tapping) instead of the original ten.
Limitations and Future Work
While the results are promising, the study was single-centered with a relatively small cohort. The authors plan to conduct multicenter studies to validate the models in more diverse environments. Additionally, while the algorithm is designed for low-cost RGB compatibility, future work must confirm these results using the varying frame rates and lens distortions common in consumer smartphones.
Conclusion
By leveraging indirect motions as proxies for rigidity and balance, this system enables a more complete remote motor assessment. This approach is particularly valuable for maintaining high-quality care and clinical trial continuity when physical exams are not possible.
Summary updated