T4

Introduction

Code for paper "Estimating Trustworthy Treatment Effects for Antibiotic Stewardship in Sepsis"

In this paper, we propose a novel method to estimate Trustworthy Treatment effects for Time-to-Treatment antibiotic stewardship in sepsis (T4).

We demonstrate that T4 can identify effective treatment timing with estimated trustworthy ITEs for antibiotic stewardship on two real-world datasets (AmsterdamUMCdb and MIMIC-III).

Requirement

Ubuntu16.04, python 3.6

Install pytorch 1.4

Data

Real-world data

Studied variables

The list of variables in MIMIC-III and AdmsterdamDB. There are 22 temporal covariates and 4 demographics andstatic variables. PT: Prothrombin Time; BUN: Blood Urea Nitrogen; WBC: White Blood Cells count.

Category		MIMIC-III		AmsterdamDB
		Mean	Std.	Mean	Std.
Demographics	Age	65.55	16.44	61.30	17.90
	Gender	43% Female	-	42% Female	-
	Weight	81.67	25.50	79.83	13.61
	Height	169.30	11.17	175.15	8.44
Lab test	Anion gap	13.35	3.80	8.70	4.62
	Bicarbonate	25.65	5.27	25.63	6.35
	Bilirubin	3.36	6.41	3.15	6.85
	Creatinine	1.50	1.46	1.28	1.03
	Chloride	104.00	6.60	108.60	46.31
	Glucose	134.00	66.83	133.9	45.74
	Hematocrit	29.96	5.13	38.98	1.67
	Hemoglobin	10.09	1.79	12.57	1.64
	Lactate	2.44	2.14	2.40	2.95
	Platelet	235.05	155.28	220.82	171.65
	Potassium	4.08	0.63	5.58	602.56
	PT	17.76	8.95	1.59	10.12
	Sodium	138.84	5.32	140.88	43.45
	BUN	29.85	23.54	14.15	9.80
	WBC	11.23	7.64	14.56	11.80
Vital signs	Heart Rate	87.81	18.30	92.70	23.65
	SysBP	120.92	23.28	126.05	139.59
	DiasBP	61.41	14.55	60.77	31.11
	Meanbp	78.70	16.88	82.12	47.34
	Respratory	20.48	5.90	21.99	7.71
	Temperature	36.96	0.85	36.73	21.14
	SpO2	97.00	3.27	96.09	7.43

Studied antibiotics

The list of antibiotics in MIMIC-III Dataset. There are 18 kinds of antibiotics in total.

Category	Name
Antibiotic	Cefazolin, Cefepime, Ceftazidime, Ciprofloxacin, Clindamycin, Erythromycin, Gentamicin, Levofloxacin, Metronidazole, Moxifloxacin, Piperacillin, Rifampin, Tobramycin, Vancomycin, Amikacin, Ampicillin, Azithromycin, Aztreonam

Synthetic data

Fully synthetic data

mkdir -p data
python simulation/gen_synthetic.py

Semi-synthetic data based on MIMIC-III

# put preprocessed MIMIC-III data into folder "data/"
python simulation/gen_synthetic_mimic.py

Train T4

Running example

bash run.sh 3 # number of follow-up steps

Results

Mortality rate comparison of two datasets. The results show that the mortality rate of patients who receive the antibiotics at the time werecommend is significantly lower than the patients who donot, indicating that our model offers effective timings of antibiotic administration that help to reduce the mortality rate.