ICD Code Tagging

This project is an attempt to implement the task of ICD code tagging with modern machine learning techniques.

Attribution

Much of this project was inspired by the following papers:

• An Empirical Evaluation of Deep Learning for ICD-9 CodeAssignment using MIMIC-III Clinical Notes
• Tagging Patient Notes With ICD-9 Codes

Additionally, ALL code in the icd9/ folder was directly cloned from this repository. This code is used for navigating the ICD9 code hierarchy via python objects. It was not available from a package manager, so it was incorporated directly to the project.

Project structure

The scripts/ directory contains python files with code for frontloading tasks that are cumbersome/expensive.

The build.py file acts as the controller for the above tasks, accomodating command-line interaction. This file can also be executed through the Makefile as described in the below section.

The report.ipynb file contains all code for exploratory/evaluation visualizations.

Reproducing code

In order to run any of this code, you must first obtain access to MIMIC-III and deploy the dataset to your AWS account.

Run pipenv install to install all dependencies.

You should create a .env file in the project root. AWS credentials should be stored here in the following variables:

• ACCESS_KEY
• SECRET_KEY
• S3_DIR
• REGION_NAME

Prerprocessing and model training can be conducted through the Makefile as follows:

• make preprocess to conduct all preprocessing steps.
• make split to split the data into train/test data
• make train to train models

Once this is all complete, you should be able to execute all cells in the report.ipynb.