Vision-based estimation of MDS-UPDRS scores for quantifying Parkinson's disease tremor severity

guannan ChenOpenReview 2025 · OpenReview

TL;DR

This paper introduces a novel video-based system, the Global Temporal-difference Shift Network (GTSN), to automatically estimate Parkinson's Disease (PD) tremor severity using the MDS-UPDRS scale. By leveraging Eulerian Video Magnification (EVM) to enhance subtle movements and a new Global Shift Module (GSM) for capturing global temporal context, the system achieves high accuracy in quantifying tremors across different body parts, offering a less burdensome alternative to manual clinical assessments.

What Problem Does This Paper Address?

Parkinson's Disease (PD) is a neurodegenerative disorder characterized by motor symptoms, with tremors being a key indicator. Clinicians assess tremor severity using the Movement Disorder Society-Unified Parkinson's Disease Rating Scale (MDS-UPDRS). However, this expert-led assessment is time-consuming, labor-intensive, and can be subjective, posing challenges for widespread and frequent monitoring, especially with an aging global population. Existing automated methods often rely on wearable sensors, which can be cumbersome, affect tremor performance, and are expensive. This paper aims to bridge this gap by developing a non-intrusive, video-based system that can accurately quantify PD tremor severity, making assessments more accessible and objective.

Methods

The proposed approach involves several key components:

  • Eulerian Video Magnification (EVM): This technique is used as a pre-processing step to amplify subtle, often imperceptible, tremors in the video. By enhancing these subtle movements, EVM makes them more detectable for the subsequent analysis.
  • Global Temporal-difference Shift Network (GTSN): This is the core deep learning model. It's built upon a ResNet-50 backbone and incorporates novel modules:
  • Temporal Difference Module: This module focuses on capturing the dynamic changes between consecutive video frames. It stacks the current optical flow with the inter-frame difference to better highlight the subtle motion patterns characteristic of tremors.
  • Global Shift Module (GSM): Recognizing that tremor assessment is a global process, this module is designed to allow each temporal segment of the video to incorporate features from all other temporal segments, not just adjacent ones. This is achieved by shifting and recombining feature channels, enabling the model to learn the overall temporal context of the tremor.
  • Video Sampling: To manage computational load while preserving crucial information, a sampling strategy is employed that adheres to the Nyquist limits, ensuring that the video frame rate is at least twice the maximum tremor frequency.
  • Body Part Segmentation: To analyze tremors from specific body parts (hands, legs, jaw) as required by MDS-UPDRS, the researchers utilized OpenPose for automatic human pose estimation, allowing them to extract relevant video segments.

Datasets

The study utilized a dataset of 130 unrelated PD patients recruited between 2019 and 2021. The diagnosis of PD was based on the UK Brain Bank diagnostic criteria. Videos were recorded during the "off-medication" state and evaluated by three trained physicians, with the ground truth determined by a majority vote. The dataset included videos for rest tremors (hands, legs, jaw) and postural tremors (hands), with specific numbers of valid videos for each body part and score category detailed in Tables 1, 2, and 3. Data augmentation techniques, such as mirroring, were applied to expand the dataset.

Key Insights and Ideas

  • EVM is Crucial for Subtle Tremors: The paper strongly demonstrates that EVM pre-processing significantly enhances the model's ability to detect and quantify subtle tremors, leading to improved accuracy, especially for distinguishing between scores of 0 and non-0.
  • Global Context Matters for Tremor Assessment: Unlike many action recognition tasks that focus on local temporal dynamics, PD tremor assessment benefits from a global understanding of temporal changes. The proposed GSM module effectively captures this global context, outperforming simpler temporal shift modules.
  • Optical Flow Outperforms RGB: The study found that using optical flow as input for the network yielded better results than using raw RGB images, highlighting that motion information is more critical for tremor detection than static visual appearance.
  • Non-Intrusive Assessment is Feasible and Effective: The research validates the potential of video-based analysis as a viable, non-intrusive method for PD tremor assessment, offering a practical alternative to wearable sensors.

Key Results

The proposed GTSN model, with EVM pre-processing, achieved impressive accuracy:

  • Rest Tremors:
    • Hands: 90.6% accuracy (multiple classification), 94.8% (binary classification)
    • Legs: 85.9% accuracy (multiple classification), 96.9% (binary classification)
    • Jaw: 89.0% accuracy (multiple classification), 95.3% (binary classification)
  • Postural Tremors:
    • Hands: 84.9% accuracy (multiple classification), 93.7% (binary classification)

The study also showed that the GSM module significantly improved performance compared to the Temporal Shift Module (TSM). The micro-average AUC for rest tremors was consistently above 0.92 across all body parts, and for postural tremors, it reached 0.93.

Limitations and Future Work

  • Data Scarcity for Severe Tremors: The paper acknowledges a limitation due to the scarcity of videos with scores of 3 and 4 (severe tremors) in the dataset. This uneven distribution impacted the accuracy for these higher scores, although binary classification (0 vs. non-0) remained strong.
  • Body Part Independence: The current model trains separate models for each body part, which increases the overall workload. Future work could explore a unified model that handles multiple body parts simultaneously.
  • Controlled Environment: The current method requires participants to adhere to specific postures and movements for video recording, which might not reflect real-world scenarios. Future research could investigate methods for obtaining ideal videos during routine daily activities.
  • Subjectivity in Higher Scores: While the model performs well, the paper notes that distinguishing between higher tremor scores (e.g., 1, 2+) can still be challenging, partly due to the inherent empirical judgment involved in the MDS-UPDRS scoring.

The authors suggest that their method has significant potential for supplementary PD tremor assessment, reducing healthcare system burden, and could be adapted for classifying other subtle motion videos.

Summary updated

Abstract

Parkinson’s disease (PD) is a common neurodegenerative movement disorder among older individuals. As one of the typical symptoms of PD, tremor is a critical reference in the PD assessment. A widely accepted clinical approach to assessing tremors in PD is based on part III of the Movement Disorder Society-Unifjed Parkinson’s Disease Rating Scale (MDS-UPDRS). However, expert assessment of tremor is a time-consuming and laborious process that poses considerable challenges to the medical evaluation of PD. In this paper, we proposed a novel model, Global Temporal-difference Shift Network (GTSN), to estimate the MDS-UPDRS score of PD tremors based on video. The PD tremor videos were scored according to the majority vote of multiple raters. We used Eulerian Video Magnifjcation (EVM) pre-processing to enhance the representations of subtle PD tremors in the videos. To make the model better focus on the tremors in the video, we proposed a special temporal difference module, which stacks the current optical fmow to the result of inter-frame difference. The prediction scores were obtained from the Residual Networks (ResNet) embedded with a novel module, the Global Shift Module (GSM), which allowed the features of the current segment to include the global segment features. We carried out independent experiments using PD tremor videos of different body parts based on the scoring content of the MDS-UPDRS. On a fairly large dataset, our method achieved an accuracy of 90.6% for hands with rest tremors, 85.9% for tremors in the leg, and 89.0% for the jaw. An accuracy of 84.9% was obtained for postural tremors. Our study demonstrated the effectiveness of computer-assisted assessment for PD tremors based on video analysis.
BibTeX
@misc{chen2025vision,
  author = {guannan Chen},
  title = {{Vision-based estimation of MDS-UPDRS scores for quantifying Parkinson's disease tremor severity}},
  year = {2025},
  url = {https://openreview.net/forum?id=UxujN0OkBy},
  howpublished = {OpenReview},
}
Figure