weiliao97 / learning_time_series_ehr

A comprehesive investigation on how time-series EHR can encode static information

Learning_Time_Series_EHR

(The repo to extract the data is in METRE and the repo to seperate sensitive latents and non-sensitive latents is in STEER)

(to be finished)

1. Extract data from MIMIC-IV an eICU using METRE

To extract data from MIMIV-IV and eICU databases, please follow the instructions in the METRE repository. To obtain the General cohort, please run:

python main.py --database MIMIC --project_id xxx

and

 python main.py --database eICU --project_id xxx

MIMIC extraction has to be run before eICU due to eICU data normalization uses extracted MIMIC-IV data statistics.

To obtain the Sepsis_3 cohort, please run:

python main.py --database MIMIC --project_id xxx --patient_group Sepsis_3

And

python main.py --database eICU --project_id xxx --patient_group Sepsis_3

2. Run preprocessing notebook

3. Run prediction models

3.1 48h in-hospital mortality task

In IHM folder, run

python main.py --model_name TCN --num_channels 256 256 256 256 --checkpoint test 

3.2 SOFA score prediction

In SOFA folder:

No fusion Transformer model

python main.py  --model_name Transformer --checkpoint test

or

python main.py --model_name TCN --num_channels 256 256 256 256 --checkpoint test 

Fusion at I, V and VI

python main.py  --static_fusion all --num_channels 256 256 256 256 --s_param 256 256 256 0.2 --c_param 256 256 0.2 --sc_param 256 256 256 0.2 --checkpoint test

With regularization l1 or l2

python main.py --static_fusion inside --num_channels 256 256 256 256 --s_param 256 256 256 0.2 --c_param 256 256 0.2 --sc_param 256 256 256 0.2 --regularization l1 --checkpoint test 

4. Run static info prediction models

In infer_static folder, have the model pt file ready for feature extraction.

TCN model with channels [256, 256, 256, 256], sepsis_3 cohort, SOFA prediction and infer race (index: 2):

python main.py --model_path xxx --model_name TCN --num_channel 256 256 256 256 --task_name sofa --read_channels 128 --infer_ind 2