README.md

PAPER TITLE TBD

AUTHORS LIST TBD

Getting Started

Installation

1. Request raw data from authors and place in ./data/raw_data/t12-updated/
2. Install project with one of the two following methods:
   1. Using conda, install environemnt with conda env create --file environment.yml, or
   2. Using vscode and docker, open this folder in vscode and run command > Dev Containers: Reopen in container
3. Copy .env.template into .env and add your weights and biases API key and path to this directory
4. Review the settings in ./configs/config.yaml. As currently set, this config will create a model that matches our results reported in the paper.
5. Run training script with ./procan_connectome/main.py from this directory. Config settings can be overridden using command line arguments. I.e., python ./procan_connectome/main.py pipeline.rfecv=False will run the training script without rfecv feature selection.

Generating figures

1. Download run results from wandb by running command python ./procan_connectome/utils/download_wandb_run_table.py. You will need access to the wandb project to obtain our results. Please contact the authors if you'd like the raw results .csv files.
2. Run notebook ./eda/updated_results.ipynb
3. Figures will be output to .plots

Existing model checkpoints

You can review our trained models by unpickling the files found in ./data/trained_models. Note that these are pickled LOOCV_Wrapper instances. See ./procan_connectome/model_training/loocv_wrapper.py for more details on this classes attributes and member functions.

Plots

EDA

Participant count vs. clinical label

Top 10 raw feature histograms

Data Preprocessing

Top 10 features histograms after transforms

Results summary

Confusion Matrix

Top 10 features by mean decrease in gini impurity