The generation folder contains the GPT-2 generation technique for i2b2 and MIMIC-III dataset. It also incorporates the differential privacy technique from Tensorflow Privacy GenerateText.py and GenerateText_i2b2.py runs the GPT generators. The 'gpt_2_simple' folder contains the DP codes included to GPT-2
Tensorflow Privacy needs to be installed prior to running the generation. The parameters for running GPT is included in GenerateText.py (or i2b2), change the input params in line 27. For example, optimizer is currently set as 'dpadam' which is the differentially private adam with noise multiplier 5.0. The epsilon can be calculated using Tensorflow Privacy package
The data from MIMIC for certain ICD9 codes ('4019', '4280', '42731', '41401', '5849', '25000', '51881', '5990') needs to be kept in seperate folder. The preprocess original file will check for these codes in a folder and preprocess the data accordingly. The input and output data both have 'ICD9 Code' at the beginning of each document. This is used in the classifiers as well.
Utility folder contains the code that tests the three utility metrics used in the paper. Adding the run parameters or settings follow BERT classifiers. To change the input check this function.
This is an academic project and the paper is under review. Contact azizmma@cs.umanitoba.ca for details