Parkinson’s Personalized Ensemble

Adaptive Deep Learning Models for Parkinson’s Diagnosis using Smartwatch Sensor Data

---

Overview

This project builds and evaluates adaptive deep learning models (CNN, LSTM) and ensemble classifiers to detect movement disorders—specifically Parkinson’s Disease (PD)—using smartwatch sensor data.

It leverages the PADS dataset (PhysioNet) and focuses on real-time task classification, condition prediction, and future personalization pipelines.

---

Objectives

• Use smartwatch accelerometer + gyroscope data from 469 patients
• Build baseline XGBoost + deep learning (CNN, LSTM) models
• Integrate patient-specific personalization via fine-tuning
• Benchmark classification performance across models
• Prepare for publication and AI competitions (e.g. Microsoft Imagine Cup)

---

Repository Structure

parkinsons-personalized-ensemble/ │ ├── data/ │ ├── raw/ # Raw .bin and JSON files (3 patients only) │ └── processed/ # Pre-extracted window data (.npy files) │ ├── notebooks/ │ ├── data_preprocessing.ipynb # Windowing and feature extraction │ ├── train_base_models.ipynb # XGBoost classifier │ └── train_deep_models.ipynb # CNN & LSTM models │ ├── paper/ │ ├── paper_draft.tex # Research draft (WIP) │ └── figures/ # Plots and visualizations │ ├── requirements.txt ├── README.md └── LICENSE

---

Dataset: PADS (Parkinson’s Disease Smartwatch)

• Source: University Hospital Münster (Germany), via PhysioNet
• Subjects: 469 individuals
• Sensors: Apple Watch Series 4 (accelerometer + gyroscope)
• Tasks: 11 guided neurological movements
• Classes: Parkinson’s Disease, Other Disorders, Healthy Controls

For full access: https://physionet.org/content/parkinsons-sensor/1.0.0/

---

Setup

1. Clone repo:

   git clone https://github.com/sonnymarmon/parkinsons-personalized-ensemble.git
   cd parkinsons-personalized-ensemble
   

Authors

Sonny Marmon – Data preprocessing, XGBoost, deep learning baseline

Harshith Guduru – Model tuning, condition classification, personalization pipeline

License

This repository is licensed under the MIT License. See LICENSE for details.

Citation

If you use this project, please cite the source dataset:

Varghese, J. et al. (2024). PADS - Parkinson's Disease Smartwatch dataset. PhysioNet. https://doi.org/10.13026/m0w9-zx22 Goldberger, A. et al. (2000). PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals. Circulation. 101(23), e215–e220.

Status

Project is under active development for submission to:

Academic publication (IEEE JBHI / arXiv)

---

Next Step:

• Copy the full markdown above and replace your GitHub README.md contents
• Commit with: