TL;DR
This paper introduces a novel machine learning approach to automatically assess Parkinson's Disease (PD) severity using video of finger tapping. By extracting features that directly align with clinical guidelines (UPDRS), their model achieves high accuracy, offering a more objective, efficient, and potentially more accessible alternative to manual clinical assessments.
What Problem Does This Paper Address?
Parkinson's Disease (PD) is a neurodegenerative disorder characterized by motor symptoms, with bradykinesia (slowness of movement) being a key diagnostic feature. The current gold standard for assessing PD severity is the Unified Parkinson's Disease Rating Scale (UPDRS), particularly its motor examination (Part III). A crucial task within this is finger tapping, used to gauge bradykinesia. However, manual UPDRS scoring is time-consuming, prone to inter-rater variability, and can be subjective. While wearable sensors and AI have shown promise, concerns about patient compliance and cost persist. Video-based analysis offers a non-contact alternative, but existing methods often rely on abstract features that lack direct clinical interpretability. This paper aims to bridge this gap by developing an automated system that uses clinically relevant features extracted from finger tapping videos to predict UPDRS scores, enhancing objectivity and explainability.
Methods
The proposed framework involves a three-cycle process: data, model learning, and deployment.
- Data Cycle: This includes data collection from diverse sources, preprocessing (upscaling, denoising, frame interpolation) to enhance video quality, and data annotation by domain experts (Movement Disorder Specialists) to provide UPDRS scores. A dual-annotator system with inter-annotator agreement analysis was used to ensure annotation quality.
- Model Learning Cycle:
- Hand Pose Estimation: The latest version of MediaPipe was used to extract 3D hand keypoints, generating time-series data of joint positions.
- Feature Extraction: A total of 15 features were extracted, including two demographic features (age, sex) and 13 kinematic features. These kinematic features were carefully designed to align with UPDRS guidelines for finger tapping, focusing on amplitude, velocity, frequency, and "halts/hesitations." Techniques like Empirical Mode Decomposition (EMD) and piecewise linear regression were employed for feature extraction.
- Machine Learning: A Decision Tree classifier was chosen for its interpretability and ability to handle ordinal UPDRS scores, aligning with the explicit criteria of the UPDRS.
- Algorithm Deployment Cycle: This involves model experimentation, updating with new data, and deployment for practical use.
Datasets
The study utilized a dataset of 75 single-view RGB videos from 50 Parkinson's Disease patients. These videos were collected in two settings: a laboratory environment (20 subjects) and a clinical environment (30 subjects). The dataset was split into a training set (60 videos) and a testing set (15 videos). The videos were captured using various cameras, including eight high-resolution cameras in the lab and an OAK-D camera in the clinic. All participants were diagnosed with PD and evaluated using the MDS-UPDRS. The raw data was manually inspected and excluded if it exhibited poor visibility.
Key Insights and Ideas
- Clinical Alignment is Key: The most significant insight is that features directly derived from UPDRS guidelines (amplitude, velocity, decrement, halts/hesitations) lead to a more interpretable and accurate model than abstract features. This "clinically informed" approach is crucial for clinical adoption.
- Decision Trees for Explainability: The choice of a Decision Tree classifier was deliberate. It allows for a clear visualization of decision-making pathways, making the model's predictions understandable to clinicians, which is a major advantage over black-box models.
- Demographics Matter (but not always in the model): While age and sex were included as potential features due to their influence on PD progression, they did not feature prominently in the final Decision Tree model. This suggests that the motor performance itself, as captured by kinematic features, is the primary driver of UPDRS scores, aligning with the UPDRS's focus on observable motor function.
- "Halts and Hesitations" Less Impactful: The study found that "halts and hesitations" had minimal impact on the model's predictions. This is attributed to the lower prevalence of upper limb freezing compared to gait freezing in PD and potentially a need for more refined definitions of these phenomena in upper limb tasks.
Key Results
The proposed Decision Tree model achieved a high accuracy rate:
- Accuracy(t1): 80% (exact match of predicted vs. true UPDRS score)
- Accuracy(t2): 93.3% (prediction within +/- 1 UPDRS score)
When compared to other methods, including AEMPD-FCN, FTTST, and baseline models using SVM and Random Forest with 15 features, the proposed F15-DT (15 features, Decision Tree) model outperformed them across various metrics, including Precision, Recall, and F1-score. For instance, its F1-score was 0.797, significantly higher than others. The confusion matrix also indicated good performance across most UPDRS score categories.
Limitations and Future Work
- Dataset Size: The authors acknowledge that the dataset of 75 videos, while yielding promising results, is relatively small and may not fully represent the broad spectrum of PD symptoms, particularly pronounced halts or hesitation phenomena. A larger dataset is needed for better generalizability.
- Scope of Motor Tasks: The study focused solely on finger tapping. Future work will expand to include other UPDRS motor tasks (e.g., hand movements, leg agility) for a more comprehensive assessment.
- Clinical Environment Realism: The study was conducted in controlled clinical environments. Future research should explore the model's robustness in more varied real-world conditions, including different lighting, patient positioning, and camera quality.
- Automated Quality Control: The manual exclusion of low-quality videos needs to be replaced by automated algorithms capable of distinguishing video quality.
- Refining "Halts and Hesitations": Further investigation into more refined definitions and time-based thresholds for "halts and hesitations" in upper limb movements is suggested.
- Ethical Considerations: The authors emphasize the importance of a multidisciplinary approach to address ethical aspects of patient privacy and data security as the technology moves towards real-world applications.
Summary updated