This repository contains the code implemented for the Master's Thesis Project "Generative Approach of Data Augmentation for Pre-Trained Clinical NLP System".

Master's Degree: "Linguistics: Text Mining", VU Amsterdam
Thesis author: Cecilia Kuan
Supervisor: Dr. Piek Vossen

Overview
This is a part of the project A-PROOF, an ongoing collaboration between AUMC and CLTL.

This thesis studies the effect of data augmentation and data sampling on improving the performance of a system classifying patients' functioning using EHRs.

Data and Class Labels
Data sets used in this thesis include 3 real data sets from previous researches, and 1 synthetic data set generated for this thesis. Previous researches use 9 class labels; 9 ICF categories are used for the A-RPOOF porject.

In this thesis, a dual-classifiers approach is used to include the 10th class, "None", to represent negative samples, as shown in the last row of the table below:

The real data sets consist of clinical notes from Electronic Health Records (EHRs) in Dutch. Due to privacy constraints, these data cannot be released. Synthetic data set can be found in the data folder.

Experiments and Evaluations
Experiments are conducted using training sets with different sampling of data, and different fine-tuning model; performance are evaluated using Precision, Recall, and F1 Scores.

Models

• Model used for synthetic data generation is OpenAI's API with model GPT-3.5;
• Models used for fine-tuning are MedRoBERTa.nl and the project's initial system, A-PROOF ICF-domains Classification created by Jenia Kim.

Cecilia_Kuan_data_generation
└───clf_domains
└───data
└───data_analysis
└───data_generation
└───data_data_process
└───ml_evaluation
└───models
└───nb_data_analysis
└───tools
└───utils
│   .gitignore
│   LICENSE
│   README.md
│   requirements.tx

• /clf_domains: scripts for training and evaluating a multi-label classification model that detects the 9 ICF domains.
• /data: data sets that can be shared
• /data_analysis: notebooks to generate statistics for corpus analysis
• /data_generation: script and notebook for generating synthetic data
• /data_process: scripts for various data processing tasks, incl. processing of raw data, processing annotations, data prep for the machine learning pipeline etc.
• /ml_evaluation: notebooks for evaluation of the machine learning models.
• /models: model files that can be shared.
• /tools: notebooks for processing outputs of dual-classifiers approach
• /utils: general helper functions used throughout the repo.

For descriptions of files within each folder, please refer to the READMEs in the individual folders.

The required Python 3.10 packages for running the code contained in this repository can be found in requirements.txt file. It is recommended to create a virtual environment with conda (you need to have Anaconda or Miniconda installed).

Some code in this repository are adapted and modified from A-PROOF repository https://github.com/cltl/a-proof-zonmw.

Code files with no changes from the original files are not included in descriptions.

Text Mining, clinical NLP, medical domain, NLP, deep Learning, Trans- formers, generative models, GPT, synthetic data, data augmentation, simpletransform- ers, MedRoBERTa.nl.