171 - Novel Use of Natural Language Processing and Artificial Intelligence to Examine Risk Factors for and Treatment of Central Venous Catheter-Related Thrombosis in a Neonatal Intensive Care Unit

Publication Number: 2165.171

Molly C. Easterlin, CHLA / University of Southern California, Santa Monica, CA, United States; Philippe Friedlich, Keck School of Medicine of the University of Southern California, Los Angeles, CO, United States; Shahab Noori, Children’s Hospital Los Angeles, Los Angeles, CA, United States; Joseph Miller, Childrens hospital los angeles, Los Angeles, CA, United States; Julie Jaffray, Rady Children's Hospital San Diego, San Diego, CA, United States; Ashwini Lakshmanan, Kaiser Permanente Bernard J. Tyson School of Medicine, Pasadena, CA, United States; Mohit Mehra, Children's Hospital Los Angeles, Thousand Oaks, CA, United States; Xinwen Li, Children's Hospital Los Angeles, South Pasadena, CA, United States

Background: Central venous catheter-related thrombosis (CVC-RT) is associated with increased morbidity, mortality, length of hospital stay, and healthcare costs. Neonatal-specific protocols do not exist for risk stratification, prevention, or management of CVC-RT due to previous inability to efficiently extract clinical data.

Objective: To use natural language processing (NLP) and a large language model (LLM) to 1) retrospectively examine associations between candidate risk factors and CVC-RT; and 2) analyze rates and characteristics of anticoagulation for CVC-RT.

Design/Methods: We included all patients admitted to a single NICU between 1/2017-9/2024 with a CVC. Using NLP of nursing documentation and clinical notes we extracted candidate CVC-RT risk factors: demographics (sex, gestational age, weight), CVC characteristics (type, lumens, size, line days), placement factors (site, vessel size, ultrasound guidance). To identify CVC-RT, we used a LLM, Qwen 2.5, on internal servers with a custom prompt, to identify thrombosis from ultrasound reports. We then assigned all thrombosis events to a potentially causative CVC. Treatment was identified by anticoagulant medication (enoxaparin) administration. We performed bivariate analyses using chi square tests, as well as univariate and multivariable generalized linear mixed effects regressions to account for fixed and random effects.

Results: Our data contain 4,203 CVCs placed in 2,277 patients. 58% of patients were male, 51% were term. At the time of CVC placement, median age was 18 days and weight 3.14 kg. Of CVCs, 68% were PICCs and 47% double lumen (Table 1).

CVC-RT occurred in 6.9% of CVCs (290/4203) and 9.9% (226/2277) of patients. 77% of thromboses occurred in deep vessels; 23% in superficial. In univariate regression analyses the following were associated with significantly increased risk of CVC-RT: younger age, PICC vs other types of CVC, longer CVC duration, smaller CVC size, lower extremity location, smaller vessel size, and larger CVC size to vessel ratio (Table 2). In multivariable regression for PICC lines, younger age, longer CVC duration, and lower extremity location were associated with thrombosis.

Of all CVC-RT, 49% of CVCs (141/290) and 46% (104/226) of patients were treated. In bivariate analyses, CVCs that received treatment were more likely to be PICCs, upper extremity, in a smaller vessel, and have a greater CVC size to vessel ratio (Table 3).

Conclusion(s): We identified risk factors for CVC-RT that could be used to inform development of a CVC-RT risk prediction tool that could help guide vascular access decisions in the NICU.

Table 1. Characteristics of Patients, Central Venous Catheters, and Placement Factors in Overall Population and by Thrombosis vs No Thrombosis


Table 2. Results from Univariate Generalized Linear Mixed Effects Regressions on Central Venous Catheter-Related Thrombosis


Table 3. Characteristics of Thromboses Treated vs. Not Treated with Enoxaparin