This is an attempt to reproduce the paper "A Novel Deep Similarity Learning Approach to Electronic Health Records Data"

V. Gupta, S. Sachdeva and S. Bhalla, ”A Novel Deep Similarity Learning Approach
to Electronic Health Records Data,” in IEEEAccess, vol. 8, pp. 209278-209295, 2020, 
doi:10.1109/ACCESS.2020.3037710.

All the code is in Jupyter notebooks. It is important to setup a virtal environement and jupyter kernel

This project requires Python 3.7+

Setup a virtual environment using the following commands

python3 -m venv ~/.venv

source ~/.venv/bin/activate pip install --upgrade pip

All the dependencies are specified in requirements.txt file. Install dependencies using

pip install -r requirements.txt

python -m ipykernel install --name=my-kernel

jupyter notebook

Open the notebooks and choose kernel "my-kernel" to run the notebooks.

The paper uses OpenEHR data which is a benchmark data available at http://www.lampada.uerj.br/en/orbda/

Data can be imported to PostgreSQL database using the scripts provided on the webiste. For this work, nephrology data was downloaded from the database and saved to csv. A sample of saved data is available under 'raw-data' folder.

• data-nephrology-s.csv.gz - Sample of ORBDA Nephrology data. This is just a 25% of data. Full data file is too big to store here. Complete data can be downloaded from orbda website. For this work, pre-processed data is available under "processed-data".

• For reusing data with different models, preprocessed data is saved under prcocessed-data folder.
  • data-interpolated.csv.gz - Pre-processed and interpolated data
  • data-normlized.csv.gz - Normalized data
  • p_x.pt.gz - Patient data with multiple visits grouped. Size (18432, 40, 42). Visits are clipped to 40 with 42 features. Uncompress this before using with the model
  • p_y.pt.gz - Target label for the patient. Visits are sorted and last label is used. Patient can have more than one label across all the visits. The dataset is specifically for chronic kidney failures. So, visits are sorted by date and picking the last one to represent the patient. Uncompress this before using with the model
  • p_y_oh.pt.gz - One hot encoded data for the target lables. Uncompress this before using with the model

• baseline-ml.ipynb - Baseline ml models using the following algorithms
  • Logistic Regression
  • Naive Bayes
  • KNN
  • Decision Trees
• baseline-cnn.ipynb - Baseline CNN model for muti-class classification
• baseline-rnn.ipynb - Baseline RNN model for muti-class classification
• baseline-mlp.ipynb - Baseline MLP model for muti-class classification
• baseline-cnn-cosine-euclidean-manhattan.ipynb - Baseline cnn model for patient similarity using cosine similarity, euclidean and manhattan distance.

All the baseline CNN models are using 2d convolution

• cnn-softmax.ipynb - Patient similarity using CNN softmax.
• cnn-softmax-balanced-data.ipynb - Patient similarity using CNN softmax using a balanced dataset.