title	PD Screening Demo
emoji	🧠
colorFrom	blue
colorTo	red
sdk	docker
app_port	8501
app_file	app.py
pinned	false

Parkinson's Disease Multimodal Screening — Interactive Research Demo

⚠️ Research demonstration only. Not a clinical tool.

Interactive Streamlit app demonstrating the attention-based multimodal fusion framework from:

"Attention-Based Multimodal Fusion of Voice and Gait for Parkinson's Disease Detection"
Aakriti Jain, Ujjawal Gaur, Pragya Singh — ICESAIA 2026 (Under Review, IEEE)

Live Demo →

---

What This App Does

Page	Description
🧠 About the Research	Paper summary, architecture, performance comparison
🎙️ Voice-Only Demo	Run inference using 22 UCI acoustic features (CSV or sliders)
🚶 Gait-Only Demo	Run inference on PhysioNet VGRF signals (upload or sample)
🔀 Multimodal Fusion	Full attention-based fusion with attention weight visualization + dropout simulation

The signature feature is the per-sample attention weight visualization — showing how the model dynamically distributes focus between voice and gait for each individual input.

---

Key Results (from the paper)

Method	Accuracy	F1	AUC-ROC
Voice-Only	86.67%	0.967	0.977
Gait-Only	89.13%	0.927	0.991
Concatenation Fusion	90.28%	0.932	0.957
Attention Fusion (Ours)	92.26%	0.942	0.986

---

Project Structure

pd-screening-demo/
├── app.py                    # Main Streamlit app (4 pages)
├── models/
│   ├── architecture.py       # Model definitions (matches paper exactly)
│   ├── __init__.py
│   ├── voice_encoder.pth     # ← NOT in repo, copy manually
│   ├── gait_encoder.pth      # ← NOT in repo, copy manually
│   ├── attention_module.pth  # ← NOT in repo, copy manually
│   ├── classifier_head.pth   # ← NOT in repo, copy manually
│   ├── scaler_voice.pkl      # ← NOT in repo, copy manually
│   └── scaler_gait.pkl       # ← NOT in repo, copy manually
├── utils/
│   ├── model_loader.py       # Cached weight loading
│   ├── preprocessing.py      # Voice + gait preprocessing pipelines
│   ├── inference.py          # Model forward pass wrappers
│   └── ui_components.py      # Plotly charts + result cards
├── samples/
│   ├── sample_pd_gait.npy    # Optional: public PhysioNet test sample
│   └── sample_healthy_gait.npy
├── export_weights.py         # Run in private training repo to export weights
├── export_samples.py         # Run in private training repo to export samples
├── requirements.txt
├── .gitignore                # Keeps .pth and .pkl out of git
└── README.md

---

Setup (Local)

git clone https://github.com/aacritea/pd-screening-demo
cd pd-screening-demo
pip install -r requirements.txt

Adding Model Weights (required for live inference)

The trained weights live in your private training repository. Use the provided export script there:

# Inside your PRIVATE training repo:
python export_weights.py \
    --checkpoint path/to/best_checkpoint.pt \
    --out /path/to/pd-screening-demo/models

Then optionally export sample files:

python export_samples.py \
    --out /path/to/pd-screening-demo/samples

The weight files are in .gitignore and will never be committed to this repo.

Running

streamlit run app.py

The app runs in demo mode if weight files are absent — showing placeholder outputs so the UI is always functional.

---

Deploy to Hugging Face Spaces

1. Create a new Space at huggingface.co/spaces

   • SDK: Streamlit
   • Visibility: Public

2. Push this repo:

   git remote add hf https://huggingface.co/spaces/YOUR_HF_USERNAME/pd-screening-demo
   git push hf main

3. Upload weights via the HF web UI or CLI (not via git):

   pip install huggingface_hub
   python - <<'EOF'
   from huggingface_hub import HfApi
   api = HfApi()
   for f in ["voice_encoder.pth", "gait_encoder.pth",
             "attention_module.pth", "classifier_head.pth",
             "scaler_voice.pkl", "scaler_gait.pkl"]:
       api.upload_file(
           path_or_fileobj=f"models/{f}",
           path_in_repo=f"models/{f}",
           repo_id="YOUR_HF_USERNAME/pd-screening-demo",
           repo_type="space",
       )
   EOF

   HF Spaces stores these as Space files, not in git history.

---

Datasets Used (Public)

• UCI Parkinson's Voice Dataset — Little (2007), DOI: 10.24432/C59C74
• PhysioNet Gait in Parkinson's Disease — Goldberger et al. (2000)

---

Citation

If you use this demo or the underlying research:

@inproceedings{jain2026parkinson,
  title     = {Attention-Based Multimodal Fusion of Voice and Gait for Parkinson's Disease Detection},
  author    = {Jain, Aakriti and Gaur, Ujjawal and Singh, Pragya},
  booktitle = {ICESAIA 2026},
  year      = {2026},
  note      = {Under Review, IEEE}
}

---

Disclaimer

This application is a research demonstration. It is not approved for clinical use and must not be used for medical diagnosis or treatment decisions.