CS598-Automated-Icd-Coding

This is the code repository for the paper Vitor Pereira, Sérgio Matos, and José Luís Oliveira. 2018. Automated ICD-9-CM medical coding of diabetic patient's clinical reports. In Proceedings of the First International Conference on Data Science, E-learning and Information Systems (DATA '18). Association for Computing Machinery, New York, NY, USA, Article 23, 1–6. https://doi.org/10.1145/3279996.3280019 as implemented by Jaya Singh (jayas2) and Mukesh Naresh Chugani (chugani2) for CS 598: Deep Learning for Healthcare.

Dependencies

Run the following command to install python dependencies

pip3 install -r final_codes/requirements.txt 

Data Download

The paper uses the MIMIC-III clinical dataset - a large, freely-available database comprising deidentified health-related data associated with over forty thousand patients between 2001 and 2012. The access to this dataset can be requested in the Physionet portal. We use the files NOTEEVENTS.csv and DIAGNOSES_ICD.csv for training and evaluating the model.

Directory Structure

Two directories: final_codes (latest version of code files) and draft_codes (old version of files i.e before draft submission)

Preprocessing

Load the pickle raw data of MIMIC-III to execute preprocessing steps and dump the resultant dataframe as a pickle file.

python final_codes/preprocess.py 

Train/Evaluate

Load the preprocessed data from above step to start training/evaluation for one of the given nine models:

Model	Command
BoT Baseline	python final_codes/<model.py/parallel_model.py> --model bot
BoT Baseline w/ Cat 0 Dense	python final_codes/<model.py/parallel_model.py> --model bot --category
BoT Baseline w/ Cat 1 Dense	python final_codes/<model.py/parallel_model.py> --model bot --category --fullyConnected 120
BoT Baseline w/ Cat 2 Dense	python final_codes/<model.py/parallel_model.py> --model bot --category --fullyConnected 200 100
BoT Baseline w/ Cat 3 Dense	python final_codes/<model.py/parallel_model.py> --model bot --category --fullyConnected 230 170 110
CNN Baseline	python final_codes/<model.py/parallel_model.py> --model cnn
CNN Baseline w/ Cat 0 Dense	python final_codes/<model.py/parallel_model.py> --model cnn --category
CNN Baseline w/ Cat 1 Dense	python final_codes/<model.py/parallel_model.py> --model cnn --category --fullyConnected 3000
CNN Baseline w/ Cat 2 Dense	python final_codes/<model.py/parallel_model.py> --model cnn --category --fullyConnected 3000 500
CNN Baseline w/ Cat 3 Dense	python final_codes/<model.py/parallel_model.py> --model cnn --category --fullyConnected 3000 700 200
CNN 3-Conv1D	python final_codes/<model.py/parallel_model.py> --model cnn3

Performance Evaluation

Following are the results obtained while training the model for 30 epochs using our implementation:

Model	Precision	Recall	F1-score
BoT Baseline	0.71	0.23	0.30
BoT Baseline w/ Cat 0 Dense	0.71	0.23	0.30
BoT Baseline w/ Cat 1 Dense	0.77	0.26	0.32
BoT Baseline w/ Cat 2 Dense	0.85	0.25	0.33
BoT Baseline w/ Cat 3 Dense	0.82	0.26	0.34
CNN Baseline	0.62	0.36	0.42
CNN Baseline w/ Cat 0 Dense	0.64	0.36	0.42
CNN Baseline w/ Cat 1 Dense	0.48	0.34	0.38
CNN Baseline w/ Cat 2 Dense	0.66	0.35	0.41
CNN Baseline w/ Cat 3 Dense	0.66	0.35	0.41
CNN 3-Conv1D	0.76	0.36	0.49