Wearable-Based Digital Biomarkers: An LSTM-Powered Progression Index for Parkinson’s Disease Monitoring

Author: Felice Dong
Advisor: Dr. Hemant Tagare
Department: Statistics and Data Science, Yale University
Completion Date: April 30, 2025

Contents

1. Overview
2. Repository Structure
3. Quick Start
4. Data
5. Model Architecture
6. Methodology
7. References and Related Work
8. Future Directions
9. Acknowledgements

Overview

Parkinson's disease (PD) affects over 10 million peopole worldwide, making objective, accessible monitoring crucial for both clinical research and patient care. The current project addresses this by developing a data-driven progression index that:

• Leverages wearable sensor data for continous, remote monitoring of motor symptoms
• Uses LSTM networks to capture temporal dependencies in activity patterns
• Provides objective measurements that complement subjective clinical scales

Key Findings

• Visual separation achieved between PD and healthy control subjects across all evaluation splits
• Peak activity capability identified as the most discriminative feature for disease status
• Weekend activities showed higher predictive value than weekday routines
• Sorted activity features provided more robust discrimination than conventional weekday labels

Repository Structure

├── src/                           # Source code
│   ├── requirements.txt           # Python dependencies
│   └── lstm_pd_progression.py     # Main LSTM model implementation
│ 
├── data/                          # Dataset files
│   ├── enhanced_fake_data.csv     # Synthetic dataset for demonstration
│   ├── fake_data.csv              # Basic synthetic dataset
│   └── fake_data_generation.ipynb # Data generation notebook
│ 
├── docs/                          # Documentation
│   ├── dong_felice_thesis.pdf     # Complete thesis document
│   └── FD_THESIS_Poster.pdf       # Research poster
│ 
├── notebooks/                     # Analysis notebooks
│   ├── thesis_model_cleaned.ipynb # Thesis model implementation
│   └── lstm_fake_data.ipynb       # Experiments with synthetic data
└── README.md                      # This file

Quick Start

Prerequisites

• Python 3.8+
• R 4.0+ (for data preprocessing)
• Required Python packages (see src/requirements.txt)

Installation

• Clone the repository

  git clone https://github.com/felice797/wearable_data_lstm_analysis.git
  cd wearable_data_lstm_analysis
  

• Install Python dependencies

  pip install -r src/requirements.txt
  

• Run the model

  python src/lstm_pd_progression.py --data_path data/enhanced_fake_data.csv --output_dir results
  

Command Line

python src/lstm_pd_progression.py [OPTIONS]

Options:
  --data_path       Path to input CSV file (default: Data_preimp.csv)
  --output_dir      Directory to save results (default: ./results)  
  --num_splits      Number of cross-validation splits (default: 20)
  --seed            Random seed for reproducibility (default: 42)

Data

CSV with the following required columns:

• subject: Unique subject identifier
• cohort: PD or Control
• week_num: Week number
• Mon, Tue, Wed, Thu, Fri, Sat, Sun: Daily ambulatory minutes

Note: The real PPMI dataset used in the thesis cannot be shared due to data privacy aggreements. Synthetic data is provided for demonstration purposes.

Model Architecure

The LSTM-based progression index uses:

• Input Layer: 7 features (daily activity minutes)
• LSTM Layers: 2 layers with 20 hidden units each
• Output Layer: Linear transformation with normalized weights
• Custom Loss Function: Optimizes separation between PD and healthy controls with time weighting

Feature Representations

1. Conventional Weekday Features: Monday through Sunday activity levels
2. Sorted Activity Features: Daily activities ordered by intensity (most to least active)

Methodology

Data Processing Pipeline

1. Quality filtering for subjects with $\geq$ 26 weeks of data
2. Missing data interpolation for gaps $\leq$ 2 days
3. Feature transformation to create intuitive progression scale
4. Temporal alignment across subjects

Model Training

• Cross-validation with 20 independent train-test splits
• Data augmentation with Gaussian noise and time swapping
• Balanced sampling to address cohort imbalance
• Early stopping based on loss convergence

References and Related Work

This research builds upon:

1. Verily Life Sciences Study (Chen et al., 2023): Digital biomarkers detected treatment effects earlier and with smaller sample sizes than traditional clinical assessments in Lewy Body Dementia patients.
2. PPMI Database: Longitudinal study providing wearable sensor data from PD patients and healthy controls.
3. LSTM Networks: Effective for capturing long-term temporal dependencies in time series data.

Future Directions

1. Enhanced Loss Function: Incorporate elements like the UPDRS score and medication information, as well as other data captured by the Verily watch
2. Missing Data Handling: Implement masking for real-world adherence patterns
3. Data Smoothing: Reduce noise while preserving underlying patterns; investigate imputation alternatives
4. Validation-Based Training: Replace convergence-based stopping with validation metrics
5. Multi-Scale Analysis: Explore different temporal scales beyond weekly aggregation

Acknowledgements

This work was completed as a Bachelor of Science senior thesis in at Yale University. Thanks to Dr. Hemant Tagare for his continuous guidance and supervision, the Yale S&DS department for their academic support, PPMI and Verily Life Sciences for providing the dataset, and my aunt—whose courage while living with PD drove this research.