Markerless Gait Asymmetry for Parkinson’s Screening

ISEF Category: Biomedical Engineering

Subcategory: Biomechanics  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

Parkinson's disease often shows up as a small foot-drag long before tremor appears. Doctors miss it because they only see the patient for ten minutes. A laptop and YouTube clips can spot the same asymmetry frame by frame. A graph neural network learns which joints carry the signal, and SHAP tells you why it decided.

What Is It?

MediaPipe Pose is a free Google library that finds 33 body landmarks in any video. Feed it a walking clip and you get a time series of joint positions. Researchers use the same library on real clinical recordings.

A graph neural network (GNN) treats the body as a graph where joints are nodes and bones are edges. Training a GNN on the joint time series teaches it to classify Parkinson's versus healthy gait. PhysioNet hosts public gait datasets you can use.

SHAP (Shapley additive explanations) opens the GNN box. It tells you which joints and which moments in the stride contributed most to the prediction. That turns a black-box classifier into a clinical explanation.

Why This Is a Good Topic

Markerless gait analysis is an active clinical research area. The pipeline is free, the data is public, and the problem is real. You will learn pose estimation, sequence modeling on graphs, and model explainability.

Research Questions

• How does input video frame rate change classifier balanced accuracy?
• What is the effect of stride normalization on Parkinson's detection?
• Does SHAP consistently flag the same joints across subjects?
• To what extent does a GNN outperform a vanilla LSTM on this task?
• Which joint feature explains the most variance in the classifier output?
• How does camera angle affect prediction stability?
• What is the effect of training-set size on subject-wise generalization?

Basic Materials

• Laptop with a GPU (or free Colab).
• Public PhysioNet gait video datasets.
• Carefully selected open YouTube clips for additional data.
• Pen and notebook for annotation log.

Advanced Materials

• Clinical-grade motion-capture data (with permission).
• Larger GPU for full GNN training.
• IRB-light protocol for any family-recorded clips.

Software & Tools

• MediaPipe Pose: Extracts body landmarks from each frame.
• PyTorch Geometric: Builds and trains the GNN.
• SHAP library: Computes feature attributions.
• OpenCV: Standardizes video preprocessing.

Experiment Steps

1. Lock the dataset sources and document every clip's provenance.
2. Define a clean subject-wise split before any modeling.
3. Decide which joints and which stride features go into the GNN.
4. Plan controls (random-shuffle baseline, age-matched controls) that rule out trivial cues.
5. Train with cross-validation and report balanced accuracy with confidence intervals.
6. Compare SHAP-flagged joints to published Parkinson's kinematic markers.

Common Pitfalls

• Mixing the same subject across train and test, which inflates accuracy.
• Using clips with widely different camera angles without a normalization step.
• Reporting raw accuracy on imbalanced classes.
• Treating SHAP output as causal explanation.
• Trusting pose-landmark coordinates from low-light or low-resolution clips.

What Makes This Competitive

A class-level project just shows a classifier accuracy number. A competitive entry uses a strict subject-wise data split so the model never sees a person in both train and test, reports balanced accuracy and confidence intervals, and validates SHAP-flagged joints against published Parkinson's kinematics literature. A fairness slice across age and sex adds maturity.

Project Variations

• Swap the GNN for a transformer and compare.
• Use only sagittal-plane projections and test whether 2D suffices.
• Add a Huntington's class and convert the task to multi-class classification.

Learn More

• PhysioNet: Free clinical gait datasets and documentation.
• PubMed: Search markerless gait Parkinson's reviews.
• MediaPipe documentation: Open-access pose-estimation guides.
• SHAP project page: Free implementation and tutorials.
• MIT OpenCourseWare: Course 6.S898 Deep Learning for Health.