Parkinson's Disease Progression Predictor 🧠

This project was submitted to Determined AI's 2023 hackathon. Check out our submission on Devpost!

We used Determined Cloud to run our models and the AMP®-Parkinson's Disease Progression Prediction database from Kaggle.

Project Objective

Our goal is to predict the progression of Parkinson's disease in patients.

Using protein abundance data, our model predicts the severity of the symptoms a patient will face in a year. The severity of a patient's condition is measured through MSD-UPDR scores, which assess both motor and non-motor symptoms associated with Parkinson's. We chose to focus on one of these scores: UPDRS 3, which measures motor symptoms including speech, facial expression, and hand movement.

Data Sample

Figure 1: This represents the distribution of the 50 features chosen using SelectKBest feature selection, after scaling and transforming. Figure 2: Shows the expected score for UPDRS, after 12 months from the visit_id index.

Getting Started

This section covers how to run our code as a Determined AI experiment with Linux and Determined Cloud.

Determined Cloud Setup

1. Create an account on Determined Cloud
2. Create a new cluster
3. Copy the cluster URL
4. Configure the DET_MASTER environment variable using export DET_MASTER=<master ip>
5. Install the Determined CLI using pip install determined
6. Log in using det auth login

Running a Training Job

1. Clone this repository
2. Enter the repository directory
3. Run det experiment create const.yaml .

Note: You can run the training job with adaptive or distributed training using one the following commands instead:

det experiment create adaptive.yaml .
det experiment create distributed.yaml .

Evaluation Metrics and Results

We used the symmetric mean absolute percentage error (sMAPE) metric to evaluate our regression model. This accuracy metric is based on percentage errors, with lower values being more desirable.

Best Metrics

Using sMAPE, our model achieved a validation loss of 0.741316.

Reproducing Results

To reproduce these results, follow the steps under Getting Started.