Representation learning on Graphs using Jumping Knowledge (JK) networks published on ICML 2018 (paper_link). In this research work, we tried to explore the concept of JK network on medical non-imaging dataset for parkinsons disease prediction.
The goal of this practical was two fold.
- Disease prediction (2 class classification)
- Exploit rich multi-modal data
- PD is a Long-term degenerative disorder of the central nervous system.
- Symptom includes shaking, rigidity, slow movement, difficulty with walking etc.
We had used 4 different types of non-imaging data for our prediction task.
- PPMI dataset (Parkinson's Progression Markers Initiative)
- MRI imaging for feature extraction
- Non-imaging data for affinity graph construction
- Apply Autoencoder, get the bottleneck, flatten for every node/patient
- LLE for dimensionality reduction
- Embedding of 324 patients, 300 features for each
- Using screening assessment results. i.e., MOCA & UPDRS test.
- Phenotype measures, i.e., Age & Gender
- At first put GCN code library in the Main Branch. Download it from GCN Github.
- Run the script downloader.py. It will automatically download necessary files in a folder named Dataset, that needed to train the network.
- Then run the main.py file for staring the training. A folder called Visualization will be created, where you will find accuracy and loss graph plots.
- Default training will run for 10 folds, each fold having 300 epochs.
- For changing the fold size or No. of epochs, please refer to the main.py file.