UIUC CS598 DL4H Final Project 2022

Team 03

Shubhendu Bhaskar - sb59@illinois.edu

Suraj Bisht - surajb2@illinois.edu

Documentation: Final report and communication with original author is available in /documentation folder

Video Presentation: https://youtu.be/Q7Fltwmcm-k

Reimplementation for paper "Improving Clinical Outcome Predictions Using Convolution over Medical Entities with Multimodal Learning"

Original paper reference: https://arxiv.org/pdf/2011.12349.pdf

  Batuhan Bardak and Mehmet Tan. 2021. Improving clinical outcome predictions using convolution over 
  medical entities with multimodal learning. Artificial Intelligence in Medicine, 117:102112.  

Original code repo: https://github.com/tanlab/ConvolutionMedicalNer

Data

Publicly available MIMIC-III dataset used for all the experiments: https://physionet.org/content/mimiciii/1.4/

Computational Requirements

All experiements performed using Colab Pro with TPU/GPU and High-RAM configuration.

Total duration requried for entire code run and results using pre-processed MIMIC-III data: 40 GPU hours

Total duration including all the additional experments and ablations: 150 GPU hours

Local setup requriments and dependencies

Jupyter notebook

Python 3.7

Tensorflow v1

Usage

1. Clone the code to local.

https://github.com/surajbisht1809/CS598_DL4H_Project_Team03_2022Spring.git
cd CS598_DL4H_Project_Team03_2022Spring

2. Run MIMIC-Extract Pipeline as explained in https://github.com/MLforHealth/MIMIC_Extract.
   Optionally download all_hourly_data.h5 from GCP at https://console.cloud.google.com/storage/browser/mimic_extract

   Pre-requisite: physionet access and GCP account

3. Copy the output file of MIMIC-Extract Pipeline named all_hourly_data.h5 to data folder.

4. Run 01-Extract-Timseries-Features_LoS.ipnyb to extract first 24 hours timeseries features from MIMIC-Extract raw data.

   Execution time: 10 mins

   Added los > 5 data points

   experimented data extraction for first 48 hours

5. Copy the ADMISSIONS.csv, NOTEEVENTS.csv, ICUSTAYS.csv files into data folder.

6. Run 02-Select-SubClinicalNotes_LoS.ipynb to select subnotes based on criteria from all MIMIC-III Notes.

   Execution time: 5 mins

   Output: sub_notes

7. Run 03-Prprocess-Clinical-Notes_LoS.ipnyb to prepocessing notes.

   Execution time: 15 mins

   Output: preprocessed_notes

8. Run 04-Apply-med7-on-Clinical-Notes_LoS.ipynb to extract medical entities.

   Execution time: 16-20 hrs - Execution is based on for loop hence most time-consuming

   Output: ner_df

9. Download pretrained Word2Vec & FastText embeddings into embeddings folder: https://github.com/kexinhuang12345/clinicalBERT

10. Run 05-Represent-Entities-With-Different-Embeddings_LoS.ipynb to convert medical entities into word representations.

    Execution time: 30 mins

    Output: six dictionary files for Wrod2Vec, Fasttext and Combined

11. Run 06-Create-Timeseries-Data_LoS.ipynb to prepare the timeseries data to fed through GRU / LSTM.

    Execution time: 30 mins

    Output: Training, validation and test data and ids

12. Run 07-Timeseries-Baseline_LoS.ipynb to run timeseries baseline model to predict 4 different clinical tasks.

    Timeseries base models used: LSTM and GRU

    Execution time: 4 hours

    Hyperparameters: hidden unites:128 and 256 epochs:10, model patience:5, Iteration:10

13. Run 08-Multimodal-Baseline_LoS.ipynb to run multimodal baseline to predict 4 different clinical tasks.

    Timeseries base models used: GRU with average multimodal using wordVec, fasttext and concat

    Execution time: 5 hours

    Hyperparameters: hidden unites:128 and 256 epochs:10, model patience:5, Iteration:10

14. Run 09-Proposed-Model.ipynb_LoS to run proposed model to predict 4 different clinical tasks.

    Proposed model: GRU with three layers of 1D convolution for wordVec, fasttext and concat embeddings

    Execution time: 8 hours

    Hyperparameters: hidden unites:128 and 256 epochs:10, model patience:5, Iteration:10

Results

Results files are available in /results folder, following are summarized result tables

Baseline vs Baseline with MultiModal

Best Baseline vs Proposed Model

References

Original Paper reference: https://arxiv.org/pdf/2011.12349.pdf

Original code repo: https://github.com/tanlab/ConvolutionMedicalNer

Download the MIMIC-III dataset via https://mimic.physionet.org/

MIMIC-Extract implementation: https://github.com/MLforHealth/MIMIC_Extract

med7 implementation: https://github.com/kormilitzin/med7

Download Pre-trained Word2Vec & FastText embeddings: https://github.com/kexinhuang12345/clinicalBERT

Preprocessing Script: https://github.com/kaggarwal/ClinicalNotesICU