PDVAT

This repository contains the code and supplementary materials for the paper titled Transformer-Based Transfer Learning on Self-Reported Voice Recordings for Parkinson’s Disease Diagnosis.

Project Overview

This project explores the use of advanced deep-learning techniques to diagnose Parkinson's Disease (PD) based on self-reported voice recordings. The methodology includes data collection, preprocessing, model training, and validation.

Methodology

1. Data Collection:

   • Voice recordings from participants using the mPower app.

2. Data Preprocessing:

   • Resampling, padding, and converting audio waveforms to log Mel spectrograms.

3. Model Architectures:

   • CNNs (ResNet50, DenseNet161) and Transformers (VIT, AST).

4. Training Regimens:

   • Training from scratch and transfer learning with fine-tuning.

5. Validation and Evaluation:

   • Subject-wise and record-wise data splits, evaluated by AUC and AUPRC.

Getting Started

Prerequisites

• Python 3.10+
• PyTorch Lightning
• Other dependencies listed in requirements.txt

Installation

1. Clone the repository:

   git clone https://github.com/Jetliqs/PD-VoiceDL.git

2. Navigate to the project directory:

   cd PD-VoiceDL

3. Install the dependencies:

   pip install -r requirements.txt

Usage

1. Prepare the dataset as described in the data preprocessing section.
2. Open and run the Jupyter notebook:

   jupyter notebook mmpd-voicedl.ipynb
   

Model Weights

We have already uploaded our model checkpoints here:

Model	Checkpoint
AST (30 epochs)	download