AI WALKUP: A Computer-Vision Approach to Quantifying MDS-UPDRS in Parkinson's Disease

Xiang Xiang, Zihan Zhang, Jing Ma, Yao Deng

TL;DR

AI WALKUP is a computer-vision-based smartphone application designed to objectively quantify Parkinson’s Disease (PD) motor symptoms. By leveraging MediaPipe for real-time pose estimation and automated feature engineering, the system transforms standard video recordings into quantitative metrics, aiming to replace subjective manual clinical ratings with consistent, remote-capable data.

What Problem Does This Paper Address?

Parkinson’s Disease (PD) is the second most common neurodegenerative disorder, characterized by motor symptoms like tremors, rigidity, and bradykinesia (slowness of movement). Currently, the "gold standard" for assessment is the MDS-UPDRS (Movement Disorder Society-Unified Parkinson’s Disease Rating Scale).

However, the MDS-UPDRS has significant bottlenecks:

  1. Subjectivity: Different clinicians may rate the same patient differently.
  2. Inconsistency: A patient’s performance can vary, and manual snapshots once every few months miss the full picture.
  3. High Cost/Low Efficiency: It requires in-person specialist visits, which are expensive and inaccessible for many.

The authors address the need for an objective, automated, and accessible tool that can quantify these symptoms using nothing more than a smartphone camera.

Methods

The researchers developed an end-to-end pipeline called AI WALKUP, which splits tasks between a mobile "User Side" and a "Cloud Side."

  • Pose Estimation: The system uses MediaPipe Pose (to track 33 body landmarks) and MediaPipe Hand (to track 21 hand landmarks). These models were chosen because they are optimized for real-time performance on mobile hardware.
  • Signal Design: The authors translated clinical MDS-UPDRS instructions into specific geometric signals. For example:
  • Finger Taps: Measured as the angle between the thumb and index finger.
  • Hand Movement: Calculated as the average distance between four fingertips and the wrist.
  • Leg Agility: Measured by the angle between the leg and the upper body.
  • Feature Engineering: To move beyond simple visualizations, the team used the tsfresh library. This allows them to extract hundreds of time-series features—such as autocorrelation, entropy, and Fourier transform coefficients—to capture the nuance of movement "quality" rather than just quantity.

Datasets

The study is a preliminary feasibility study. The authors collected action video data from 20 healthy individuals. Each participant performed six specific MDS-UPDRS motor tasks (e.g., finger tapping, hand opening/closing, leg agility). While the current dataset focuses on healthy controls to validate the signal extraction pipeline, the authors note that the algorithm is designed to be easily extended to PD patients in future clinical trials.

Key Insights and Ideas

  • Democratizing Diagnostics: The core idea is that the smartphone in your pocket is a sophisticated motion-capture lab. By moving the analysis to the cloud, the app remains lightweight while the heavy computational lifting (feature extraction) happens elsewhere.
  • Quantifying "Fatigue": One of the most interesting aspects of the results is the visualization of signal decay. In several tasks, the signals show that while the amplitude of movement remains steady, the frequency decreases over time (the gaps between peaks get wider). This "slowing down" is a critical clinical marker for PD that is often hard to quantify precisely by eye.
  • Beyond Peak Counting: By using tsfresh, the authors recognize that Parkinsonian motion is complex. Features like "approximate entropy" or "variation coefficients" can detect irregularities in rhythm that a human observer might miss but are mathematically distinct in a time-series.

Key Results

  • Successful Signal Extraction: The system successfully extracted clear, periodic signals for all six motor tasks from monocular (2D) smartphone video.
  • Mobile Deployment: The authors successfully deployed the "AI WALKUP" app on the Android platform, proving that the workflow (Record -> Upload -> Analyze -> Result) is viable for non-technical users.
  • Qualitative Validation: Visualizations of the extracted signals (Figs 5-8 in the paper) clearly show the ability to track maximum/minimum amplitudes and the speed of repetitions, providing a "digital twin" of the physical clinical exam.

Limitations and Future Work

  • Clinical Validation: The most significant limitation is that the current study was performed on healthy subjects. The next step is to correlate these digital features with actual MDS-UPDRS scores (0-4) assigned by neurologists to see if the AI's "score" matches the human expert.
  • Cross-Patient Variability: The paper does not yet address how different body types, lighting conditions, or camera angles might affect the feature extraction.
  • Longitudinal Analysis: The authors plan to perform "over-time" analysis to see if the system can detect the subtle progression of the disease over months or years, which would be a breakthrough for clinical drug trials.

Summary updated

Abstract

Parkinson's Disease (PD) is the second most common neurodegenerative disorder. The existing assessment method for PD is usually the Movement Disorder Society - Unified Parkinson's Disease Rating Scale (MDS-UPDRS) to assess the severity of various types of motor symptoms and disease progression. However, manual assessment suffers from high subjectivity, lack of consistency, and high cost and low efficiency of manual communication. We want to use a computer vision based solution to capture human pose images based on a camera, reconstruct and perform motion analysis using algorithms, and extract the features of the amount of motion through feature engineering. The proposed approach can be deployed on different smartphones, and the video recording and artificial intelligence analysis can be done quickly and easily through our APP.
BibTeX
@misc{xiang2024ai,
  author = {Xiang Xiang and Zihan Zhang and Jing Ma and Yao Deng},
  title = {{AI WALKUP: A Computer-Vision Approach to Quantifying MDS-UPDRS in Parkinson's Disease}},
  year = {2024},
  eprint = {2404.01654},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2404.01654},
}
Figure