Skip to content

Research project done in Fall 2019 on measuring a treatment effect with causal inference, with missing data imputation, PCA and hierarchical clustering.

Notifications You must be signed in to change notification settings

shmodal/Effect_of_a_treatment_over_trauma_patients

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Effect of a treatment over trauma patients

Traumabase®, (http://www.traumabase.eu) is a a growing database built from the collaboration between several trauma centres in France. Severe trauma represents one of the most common cause of death in patients 16-45 years of age. Several initiatives have been undertaken to provide support and tools for doctors, to help them assess their best practices and provide an efficient response to increasing health challenges.

The data is confidential and as such cannot be shared on any outside platform.

In this project, we studied the effect of tranexamic acid, a treatment administered to patients suffering heavy bleeding consecutive to a head trauma.

We used a raw dataset of initially 7000 rows and 250 features, with both categorical and continuous variables. We performed an investigation of the missing values according to their class (missing completely at random, missing at random and missing not at random). With an eye on the former, we critically imputed missing data through two different methods, random forests and factorial analysis for mixed data (FAMD), for comparison purposes.

We then ran a descriptive analysis using principal components analysis (PCA) and hierarchical clustering. We determined the most relevant variables, their correlation with cranian trauma, and found three relevant clusters of patients exhibiting similar characteristics. We validated our findings with doctors.

Afterwards, in order to measure the effect of the treatment over mortality, we had to use several causal inference techniques. Indeed, tranexamic acid is predominantly administered to patients in hemorrhagic shock with a high risk of dying.

This reveals an attribution bias and tends to skew the average treatment effect. To account for the covariates, we tried matching, inverse propensity weighting (IPW), and double robust estimators.

About

Research project done in Fall 2019 on measuring a treatment effect with causal inference, with missing data imputation, PCA and hierarchical clustering.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages