Overview

TBI Patient Endotype Clustering - using LCA to demonstrate the existence of patient endotypes in a TBI population that differ in their comorbidity profiles and clinical outcomes.

Methodology

Core data sources

• Requires MIMIC III v1.4 database to be built in SQLite format locally

1. tbi2_admit_icd.xlsx - contains all TBI patient demographic and comorbidity information (e.g., DOB, sex, comorbidity ICDs, discharge status)
   • Query found in MIMIC3 Linked TBI Desktop v2.accdb MS Access file under tbi_admit_icd query
   • Note that the .accdb file must be manually linked to the locally built SQLite MIMIC III database in order for the query to be run
2. tbi2_admit_proced.xlsx - contains neurosurgical intervention events for TBI patients
   • Full query found in sql_queries.py
3. tbi2_admit_chevents_gcs.xlsx - contains recorded GCS scores for TBI patients
   • Full query found in sql_queries.py

Data analysis pipeline

1. The core data sources are combined and processed by processing.py to generate tbi2_admit_icd_dates_nsx_gcs_elix.xlsx containing all key clinical information of interest
2. Using ICD-9 data from tbi2_admit_icd_dates_nsx_gcs_elix.xlsx, comorbidity endotypes are identified using cluster.r and appended for each patient onto the original Excel file to create tbi2_admit_icd_dates_nsx_gcs_elix_annotated.xlsx
3. Statistical analyses and visulizations are generated using data from tbi2_admit_icd_dates_nsx_gcs_elix_annotated.xlsx using the following files:
   • graph_builder.py and graph_renderer.py for comorbidity network graphs and relative risk between comorbidities
   • visulization_stats.py for conventional descriptive statistics and stratified analyses (by GCS and age)
   • cluster.r for visulization of comorbidity distributions
   • visulization_stats.r for jitter plot and ANCOVA and regression analysis for endotypes

Demo