Example:
| Name | GitHub Handle |
|---|---|
| Paulene Patrisha Pare | @PaulenePatrishaPare |
| Helen Ton | @heleeon |
| Irene Chang | @rose413 |
| Maria Lopez | @mitchizepic |
| Wenting Lu | @wlu32 |
Developed a machine learning pipeline using XGBoost and Random Forest classifiers to detect gait events and early signs of Freezing of Gait (FoG) in Parkinson’s patients using wearable sensor data. Achieved 82% accuracy and strong ROC–AUC performance, demonstrating the potential of accelerometer-based models to support objective mobility assessment. Generated insights into movement irregularity, energy patterns, and rhythmicity changes that occur during FoG-related behaviors such as Start Hesitation. Implemented a complete pipeline including signal filtering, windowing, feature engineering (time + frequency domain), TSFresh extraction, and model evaluation, addressing real-world constraints in sensor noise, class imbalance, and short-duration events.
HOW TO CLONE THE REPOSITORY
Open Terminal (Mac/Linux) or Git Bash (Windows). Run this command: git clone https://github.com/heleeon/MJF-1B-Parkinsons-Disease.git
Move into the project folder: cd MJF-1B-Parkinsons-Disease
You now have a full local copy of the repo on your machine.
HOW TO EDIT THE FILE
Use VS Code Open VS Code Go to File → Open Folder Select the cloned folder: MJF-1B-Parkinsons-Disease/ You can now edit: notebooks (.ipynb) scripts (.py) README (README.md) any other files
Use Jupyter Notebook Inside the repo: jupyter notebook Then open the notebook file you want to edit.
HOW TO SAVE AND UPDATE EDITS ON GITHUB Once you edit files locally, run: git add . git commit -m "Updated README and added new features" git push
This requires: You to be logged into GitHub You to have permission to push to the repo (if it’s your repo, yes; if not, you need to fork it first)
If you don’t have push permissions: If you get a “permission denied” error: Click Fork in GitHub Clone your version of the repo: git clone https://github.com/YOUR_USERNAME/MJF-1B-Parkinsons-Disease.git Edit locally Push changes to your fork Submit a Pull Request if you want the owner to merge your edits
This project was completed as part of the Break Through Tech AI Studio, where students partner with an industry host to solve real-world AI/ML problems. Our host company focuses on health technology and mobility assessment, providing the challenge of identifying gait disruptions from wearable sensors. Parkinson’s Freezing of Gait is a major cause of falls and loss of independence. Early and accurate detection is crucial for monitoring disease progression, improving therapies, and supporting daily mobility. This project builds a fully reproducible ML workflow capable of transforming raw accelerometer signals into predictions of clinically meaningful gait events.
Dataset Description Sources: Defog and Tdcsfog datasets Format: 3-axis accelerometer data with event labels (Normal, Walking, Turn, Start Hesitation) Size: ~7k–30k windows depending on preprocessing
Challenges: Class imbalance Variable signal quality Short-duration events Need for alignment with event timestamps
EDA Insights FoG-related events show lower movement energy, less rhythmic patterns, and higher spectral entropy. Turns exhibit larger variability in acceleration magnitude. Start Hesitation shows subtle but detectable irregularities in frequency content.
Visualizations Included Raw vs. band-pass–filtered signals Distribution plots of energy, entropy, freezing index Correlation heatmaps Event-wise scatterplots
Models Used Random Forest – strong baseline for tabular sensor features XGBoost – top-performing model across most metrics SVM – comparison baseline CNN (optional exploration) – for sequence-based modeling
Feature Engineering 800+ combined features from: Time-domain statistics Frequency-domain metrics Freezing Index Energy + entropy TSFresh automated features Cross-axis correlations
Training Setup Train/test split with stratification Evaluation metrics: Accuracy, Macro F1, ROC–AUC Baseline established using majority-class classifier
Insights FoG-associated behaviors show consistent shifts in energy, entropy, and rhythmicity. Feature importance reveals that vertical-axis variability and mid-frequency energy bands are key indicators of gait instability. ROC curves demonstrate strong class separability for Normal and Start Hesitation.
Visualizations Included Confusion matrices ROC curves (model comparison) Feature importance plots Distribution and scatter plots
Explore different models Reduce the size of the dataset Improve and develop the features Create more visuals to understand the data Optimize the model parameters Aim for higher accuracy, precision, f1, and recall scores
This project is licensed under the MIT License.
Kaggle Parkinson’s FoG datasets TSFresh Feature Extraction Library Documentation XGBoost and scikit-learn documentation
Thank you to: Harshini Donepudi, our AI Coach Seth Haney and Barbara Marebwa, our Challenge Advisors Break Through Tech AI teaching assistants for all their support in our AI Studio project!