SincPD: An Explainable Method based on Sinc Filters to Diagnose Parkinson's Disease Severity by Gait Cycle Analysis

Overview

This repository hosts the implementation of an explainable deep learning model (SincPD) for Parkinson’s Disease (PD) diagnosis and severity estimation through gait cycle analysis. Leveraging SincNet filters as adaptive bandpass filters, our approach extracts key frequency features from vertical Ground Reaction Force (vGRF) signals. Such a design promotes better interpretability compared to conventional deep neural networks.

Paper on arXiv

A preprint of this work is now available on arXiv.
Please cite this paper if you use or build upon our work.

Citation

@misc{salimibadr2025sincpdexplainablemethodbased,
      title={SincPD: An Explainable Method based on Sinc Filters to Diagnose Parkinson's Disease Severity by Gait Cycle Analysis}, 
      author={Armin Salimi-Badr and Mahan Veisi and Sadra Berangi},
      year={2025},
      eprint={2502.17463},
      archivePrefix={arXiv},
      primaryClass={eess.SP},
      url={https://arxiv.org/abs/2502.17463}
}

Key Contributions

• Adaptive Sinc Filters: Extract important frequency components from raw vGRF signals.
• Explainable AI: Pruning filters with clustering techniques (K-means + silhouette scores) to highlight significant frequency bands.
• PD Diagnosis & Severity Estimation: Achieves high accuracy for both classification and severity levels.
• Preprocessing & Optimization: Removes noise from signals and prunes redundant filters to enhance efficiency and interpretability.

Model Architecture

1. Preprocessing: Standardization and segmentation of gait cycle signals.
2. Feature Extraction: SincNet layers learn adaptive bandpass filters specific to PD vs. healthy signals.
3. Pruning & Optimization: Clustering-based approach to prune filters, ensuring a concise and interpretable model.
4. Classification: Final fully connected layers classify PD presence and severity.

Dataset

The dataset used in this study is from PhysioNet's Gait Database.

Performance

• 98.77% accuracy for PD vs. healthy classification.
• 97.22% accuracy for PD severity prediction.

Repository Structure

├── data/            # To store the gait dataset (not included by default)
├── models/          # Model definitions (SincNet + custom layers)
├── notebooks/       # Jupyter notebooks for running experiments & demos
├── results/         # Evaluation results & visualizations
├── requirements.txt # Project dependencies
└── README.md        # Project documentation (this file)

Installation & Usage

1. Clone the repository:

   git clone https://github.com/<YOUR_USER_OR_ORG>/SincPD.git
   cd SincPD

2. Install dependencies (example):

   pip install -r requirements.txt

3. Run Jupyter notebooks for experiments (under notebooks/):

   jupyter notebook

Future Work

• Pretrained Models: Providing downloadable weights for quick inference.
• Generative Modeling: Generating synthetic gait data for data augmentation.
• Domain Expansion: Adapting the approach to other signal domains (e.g., protein sequences).

Contributors

• Mahan Veisi
• Sadra Berangi
• Armin Salimi-Badr

License

The license will be determined based on journal/publisher requirements and will be updated here.

---

If you find this project useful in your research, please consider starring the repository and citing our arXiv paper. Thank you!