🚶🏻‍♀️ PB2PD — Prodromal Burden to Parkinson’s Disease — MVP

From in-PD prodromal burden to prospective screening tools.
This project extracts and explains gait-based biomechanical signatures associated with the “prodromal burden” phenotype (PB+ vs PB–) in patients already diagnosed with Parkinson’s disease.
The resulting signatures are designed to inform future screening studies on normative subjects with prodromal symptoms.

---

🔍 Overview

PB2PD analyzes lumbar IMU–derived gait features from a PD cohort stratified by prodromal burden (PB+ / PB–), where PB includes combinations of:

• REM Sleep Behavior Disorder (RBD)
• Hyposmia
• Depression
• Constipation

The aim is to identify stable biomechanical signatures, validated with explainable AI, that may guide future studies on prodromal subjects without diagnosed PD.

Key components:

• Lazy Prediction for benchmark screening
• Balanced Random Forest (with ctGAN data balancing)
• SHAP explainability (global + local)
• Surrogate clinical rules (interpretable trees)orts.

---

Webapp (Streamlit MVP)

Patient input	Risk results	PDF export

Key features:

• Single patient or batch CSV input.
• Color-coded risk bands (green/yellow/orange/red).
• Screening mode: set prevalence → PPV/NPV estimate.
• Local SHAP contributions (top-5) for clinical transparency.
• Surrogate rules (interpretable tree) for clinical validation.
• PDF export with minimal QC, thresholds, SHAP, and rules.

---

Architecture

• Notebooks (EDA, preprocessing, explainability, robustness) → /notebooks
• Models & scaler → /models
• Surrogate rules (JSON/YAML) → /artifacts
• Streamlit app → /app/app.py
• Results/metrics → /tables and /figurez

Integrated pipeline → from biomechanical analysis to clinical reports.

---

Key Techniques

• Generative data balancing → ctGAN
• Advanced explainability → SHAP, SHAPSet plot, surrogate tree with 95% CI
• Robustness analysis → cross-validation, subgroup analysis, calibration
• Decision support → Decision Curve Analysis (Net Benefit)

---

Example of PDF Report

Includes:

• Patient input
• Triad probability + risk band
• Sensitivity/Specificity @ threshold
• PPV/NPV at custom prevalence
• SHAP top-5 contributions
• Surrogate rule matched
• Minimal QC (range, missing, sex, H-Y)

---

🚀 Setup & Run

Clone the repo and run the Streamlit app:

git clone https://github.com/DanteTrb/PB2PD.git
cd PB2PD/app
pip install -r ../requirements.txt
streamlit run app.py

🧑‍🔬 Clinical Research Disclaimer

•	Research/triage use only.
•	Does not replace clinical judgment.
•	Not a certified medical device.

🏆 Credits & Vision

•	Author: Dante Trabassi, Sapienza University of Rome