Self-Supervision for Semi-Supervised Prediction of ESKD Outcome

This repository contains codes for training deep-learning networks for predicting binary outcome from various forms of input features

• FCNN: For predicting disease outcome from patient-level features (i.e. Urinary Proteomics)
• RNN: For predicting disease outcome from structure-level features within patients (i.e. Digital Image Features from glomeruli, tubules)

Data

Whole slide DN biopsies can be found at athena.rc.ufl.edu

Requirements

FCNN:

• Tensorflow 2.6.1
• NumPy
• Pandas
• Sklearn

RNN:

• Python 3.6.5
• Tensorflow 1.14.0
• NumPy
• Pandas

Usage

FCNN:

Edit the "model_ESRD_agg.py" script to properly point to data locations

data = pd.read_csv('Path_to_Data_With_Label.csv').to_numpy()
val_data = pd.read_csv('Path_to_Data_Without_Label.csv').to_numpy()
labels = pd.read_csv('Path_to_Labels.csv').to_numpy()

Then run the script:

python3 model_ESRD_agg.py

RNN:

Edit the Kfold_RNN.py script to properly point to data locations generated by codes found at: github.com/njlucare/FeatureExtraction/

txt_loc = 'path_to_features_with_label.txt'
lbl_loc = 'path_to_labels.csv'

txt_loc_v = 'path_to_features_without_label.txt'
lbl_loc = 'arbitrary_path.csv'

Then run the script:

python3 Kfold_RNN.py

Both scripts will output loss and predictions text files with SoftMax Outputs and labels

Acknowledgements

Please cite our work @