Using sequences of clinical codes for treatment effect estimation

Thumbnail Image

Files

master_Luong_Nguyen_2021.pdf (2.02 MB)

URL

Journal Title

Journal ISSN

Volume Title

Perustieteiden korkeakoulu | Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.

Authors

Date

2021-08-23

Department

Major/Subject

Computer Science

Mcode

SCI3042

Degree programme

Master’s Programme in Computer, Communication and Information Sciences

Language

en

Pages

55

Series

Abstract

Estimating causal effect from observational data has recently been gaining traction in the research community. Much of this heightened interest is owed to the increased availability of large data resources, for instance, Electronic Health Records (EHR)and social media posts, which have effectively transgressed the known limitation of randomized controlled trial (RCT) such as time and budget requirements. Embraced by the rapid development in the machine learning domain, a plethora of methods have been introduced to analyze treatment effects from observational data. Nevertheless, the majority of the proposed methods have been operated in a static setting that is not applicable to clinical environments whose patients’ states are ever-changing. Therefore, the present thesis aims to develop a novel approach utilizing clinical codes as covariates to estimate treatment effect from synthetic medical data. Two neural network (NN) models, namely Long Short-Term Memory (LSTM)-based and Transformer Encoder-based, are employed to generate a representation from raw clinical codes that is suitable for treatment effect inference. With a simple synthetic dataset, the NN models are found to be on par with the Lasso model when the treatment is equally distributed and moderately more propitious in the unbalanced case. Future work includes benchmarking the models on more complex synthetic data as well as real data. Additionally, the presence of hidden confounders in time-series data is a rather essential aspect that requires meticulous attention before evaluating the treatment effects. Thus, adjusting for the effects of hidden confounders is another potential research direction.

Description

Supervisor

Marttinen, Pekka

Thesis advisor

Kumar, Yogesh

Keywords

machine learning, deep learning, treatment effect estimation, electronic health record, clinical data, causal inference

Other note

Citation

Permanent link to this item

https://urn.fi/URN:NBN:fi:aalto-202108298501

Collections

[dipl] Perustieteiden korkeakoulu / SCI