Welcome the the BlendedICU code repository

Usage:

Setting up the BlendedICU database

The path to the BlendedICU database (version 0.3.2) should be specified in paths.json.

{
    "blendedICU": "D:/BLENDED_ICU/0.2.3/blended_data/"
}

Running the BlendedICU harmonization pipeline requires acess to AmsterdamUMCdb, eICU, HiRID and MIMIC-IV.

For more information on the BlendedICU database, check the article and GitHub repo.

Running BlendedICU

To facilitate reproduction of the results, scripts used for producing tables and figures were numbered 1 to 5.

1. split_train_test

Data is partitioned using the source database (Amsterdam, eICU, HiRID and MIMIC-IV). The same number of patients is extracted from each partition: source databases are equally represented in the extracted data. This step prepares a training, validation and test set of each partition.

2. run_experiment

This is the main part of the code. A set of experiments can be specified in config.json

The config block contains parameters that are shared between all experiments. Parameters that are not specified in this block are given default values in models/config.py.

{
    "config": {
        "task":"multitask",
        "dataset": "blended",
        "use_flat": true,
        "use_med": true,
	    "use_lab": true,
        "use_vitals": true,
        "use_resp": false,
	"percentage_test": 50,
	"alpha": 20
    },
    "experiments":{
        ...
        
    }
}

The experiment block contains parameters of each experiment. If specified in both blocks, the experiment block override parameter definition from the config block.

An experiment is defined by several runs. In a single run: a set of models is trained and evaluated with common parameters.

{
    "config": {
        ...
    },
   "experiments": {
	    "main_experiment":{
	        "models": ["tpc"], # The TPC from Rocheteau & al is used in this experiment
	        "runs":[
	             {"train_on":["mimic4"], # Training on MIMIC-IV, and keep as pretrained model
	              "percentage_trainval": 75,
	              "use_as_pretrained": true},
	             {"train_on":["amsterdam"],  # Training on AmsterdamUMC
	              "percentage_trainval": 25},
	             {"percentage_trainval_dic": {"amsterdam": 25, # Data pooling of MIMIC-IV and AmsterdamUMC
					                          "mimic4": 75}},
	             {"train_on": ["amsterdam"], # Transfer Learning from pretrained model
	              "percentage_trainval": 25,
	              "from_pretrained": true,
	              "retrain_last_layer_only": false}
	            ]
	    },
        "model_benchmark": {
	        "models": ["lstm", "transformer", "tpc"],  # Three models are run in this experiment
	        "runs":[
	            {"train_on": ["amsterdam", "hirid", "mimic4", "eicu"],
                 "percentage_trainval": 20} #There is a single run with pooled data from each source database.
                ]
        },
    }
}

3. generate_table

Produces the results tables (including appendix) in pandas and Latex format. The names of the experiments from step 2 should be filled in the Experiments constructor.

self = Experiments(
    main_experiment='main_experiment',
    model_benchmark='model_benchmark',
    dataset_benchmark='dataset_benchmark',
    dataset_benchmark_nomed='dataset_benchmark_nomed',
    training_size='training_size_study',
    )

4. extraction_table

This code produces the main table decribing the extraction of source database.

Questions ?

Feel free to open an Issue !