327 Identification of Clinically Relevant Patient Endotypes in Traumatic Brain Injury (TBI) Using Latent Class Analysis

Abstract

INTRODUCTION: Despite advances in medical knowledge, TBI still has a significant burden of disease. Better prognostic factors would allow for earlier identification of patients at risk for poor outcomes and enable targeted interventions to improve outcomes and reduce costs. METHODS: We obtained clinical information from a subset of TBI patients from the MIMIC-III dataset and codified each patient's comorbidities based on the Elixhauser comorbidity groups. Latent class analysis was performed on the patients based on their codified comorbidities to identify distinct patient endotypes. The characteristics (comorbidity distribution, mortality rate) of these patient endotypes were then visualized using Matplotlib. A Chi-squared test of homogeneity was performed to determine if mortality rates differed across patient endotypes. Relationships between the codified comorbidities were visualized using a network graph across all TBI patients and within each patient endotype. RESULTS: 2981 patients qualified under our TBI query. Average age was 58 ± 21. The network graph shows that there is only one major connected component composed of the most common comorbidities. An elbow plot was constructed to help identify 8 as the optimal number of classes to model patient endotypes. The distribution of comorbidities differs between patient endotypes. Furthermore, there was a significant difference in survival rates between the patient endotypes (p < 0.0001) with Endotype 1 having the lowest survival and Endotype 6 having the highest. Intra-endotype network analysis and comorbidity distributions suggest that Endotype 1 is a cancer-related endotype whereas Endotype 6 is a relatively "healthy" endotype with some degree of alcohol and drug abuse. CONCLUSIONS: In this study, we use LCA to demonstrate the existence of patient endotypes in a TBI population that differ in their comorbidity profiles and clinical outcomes.