DL4H-Automated-ICD-9-Coding

Author: Ryan Fogle

Overview

The project is split up into 7 parts.

• Prepare-data
• SVM Training
• Word2Vec Training
• Doc2Vec Training
• DeepLabeler Training
• DeepLabeler minus Doc2Vec Training
• DeepLabeler with Embedding Layer

All of these trainings happen sequentially, please go through one by one.

Environment setup

This environment setup uses python 3.10, a 6 core CPU, and a Nvidia 3080 10GB. You'll need to install the dependencies as well.

cd DL4H-Automated-ICD-9-Coding
pip install -r requirements.txt

Report

The report can be seen in report.pdf

Computation Times

Model Metrics

Summary of Paper

ICD-9 coding is a time-consuming task that requires a specialized skill set to provide accurate ICD-9 codings. Using the discharge summaries and ICD-9 diagnostic codes provided in the MIMIC-III dataset, the paper ”Automated ICD-9 Coding via A Deep Learning Approach” introduces a multi-label classification problem. The authors of the paper introduce a deep learning model called DeepLabeler which incorporates two models: a Word2Vec model (Mikolov et al., 2013) and a Doc2Vec model (Le and Mikolov, 2014). The paper claims DeepLabeler performs better (via an F1 score) than a traditional natural language processing model like a support vector machine (Li et al., 2019).

Summary of Findings

The results of this report do not support all of the claims made in the paper. The main premise of the paper is that by adding a deep learning architecture we can increase the micro-f1 score, my findings do not support that claim. Although, this report does support the claim that by adding the Doc2Vec vectors, the micro-f1 score increases.

References

Gensim library: https://radimrehurek.com/gensim/

sci-kit learn: https://scikit-learn.org/stable/

pytorch library: https://pytorch.org/