ruoqi-liu / dsw

DSW

Introduction

This repository contains source code for paper "Estimating Individual Treatment Effects with Time-Varying Confounders".

In this paper, we study the problem of Estimating individual treatment effects with time-varying confounders (as illustrated by a causal graph in the figure below)

We propose Deep Sequential Weighting (DSW) for estimating ITE with time-varying confounders. DSW consists of three main components: representation learning module, balancing module and prediction module.

To demonstrate the effectiveness of our framework, we conduct comprehensive experiments on synthetic, semi-synthetic and real-world EMR datasets (MIMIC-III). DSW outperforms state-of-the-art baselines in terms of PEHE and ATE.

Requirement

Ubuntu16.04, python 3.6

Install pytorch 1.4

Data preprocessing

Synthetic dataset

Simulate the all covariates, treatments and outcomes

cd simulation
python simulate_full.py

Semi-synthetic dataset

With a similar simulation process, we construct a semi-synthetic dataset based on a real-world dataset: MIMIC-III.

cd simulation
python simulate_mimic.py

Full list of covariates

We show the full list of static demographics and time-varying covariates of sepsis patients obtained from MIMIC-III.

Category	Items	Type
Demographics	age	Cont.
	gender	Binary
	race (white, black, hispanic, other)	Binary
	metastatic cancer	Binary
	diabetes	Binary
	height	Cont.
	weight	Cont.
	bmi	Cont.
Vital signs	heart rate, systolic, mean and diastolic blood pressure	Cont.
	Respiratory rate, SpO2	Cont.
	Temperatures	Cont.
Lab tests	sodium, chloride, magnesium	Cont.
	glucose, BUN, creatinine, urineoutput, GCS	Cont.
	white blood cells count, bands, C-Reactive protein	Cont.
	hemoglobin, hematocrit, aniongap	Cont.
	platelets count, PTT, PT, INR	Cont.
	bicarbonate, lactate	Cont.

DSW

Running example

python train_synthetic.py --observation_window 30 --epochs 64 --batch-size 128 --lr 1e-3

Outputs

• ITE estimation metrics: PEHE, ATE
• Factual prediction metric: RMSE

dsw's Issues

No "simulation" directory

Hi! Thanks for making the codes available. I would like to try out the model on MIMIC-III dataset, but seems the directory "simulation" containing the pre-processing codes (cd simulation; python pre_mimic.py) is not in the repo. Would it be possible to also publish this directory?

Thanks for your help!

no pre_mimic.py

Hello, I was not able to find pre_mimic.py. I was wondering if the file may have been deleted or if it is yet to be shared. Thank you.

How to train model without Y_cf

Hi Ruoqi, we are trying to apply your model to our real data problem. I know your code is for simulation data, so we have to replace your data with our real data. Since for convenience, we would like to change as few codes as possible, so we decide to replace your A, X.all, X.Static and Y with our real data. but when we are planning so, we find out that there is Y_cf in your code. So i am wondering if we can get rid of Y_cf. or more a more general question, how can we import real data and apply your code to train the model. Thanks very much to any response or hint!