Abstract Body (Do not enter title and authors here): Background:
Central venous catheter (CVC)-related venous thromboembolism (VTE) is a common complication in the pediatric cardiac intensive care unit (PCICU), with the incidence rate of 2-18%. Clinical suspicion of thrombosis arises from observations of edema, pain, tenderness, and phlebitis, and is subsequently confirmed with ultrasound. However, nearly 40% of thrombotic events are asymptomatic, and routine ultrasound surveillance is neither practical nor cost-effective. Once detected, VTE can be treated with anti-coagulation therapy. Thus, while detection and treatment are relatively straightforward, initial suspicion relies solely on clinical observation.

Research Question:
Our research aims to develop an alert system that can automatically suspect thrombosis from high-frequency central venous pressure (CVP) data, notifying the clinical team of potential VTE events. Our initial efforts suggest that most false positives generated by our alert system are due to impulsive noise. Therefore, in this work, we have developed an algorithm to quickly remove impulsive noise artifacts from CVP waveforms to reduce false positives in our VTE detection algorithm.

Methods:
Patients in the PCICU at Dell Children's Medical Center (Austin, TX), from December 2018 to January 2021, with ultrasound-confirmed VTE were included in this single-center retrospective trial. Patient-specific CVC details (type, size, and anatomical location), age, and date of insertion were recorded. CVP signals were sampled at 125 Hz using the Sickbay Platform. CVP waveforms were decomposed into a series of intrinsic mode functions (IMF). Impulsive noise was eliminated through robust measures of center and scatter of the IMFs.

Results:
Thirty pediatric patients with ultrasound-confirmed VTE were included in this pediatric cohort. Manual curation identifies 6.0% ± 3.5% of data in the four hours near catheter insertion as impulsive noise. Our algorithm can currently detect the manually curated noise with up to 90% sensitivity and 95% specificity, with a large portion of the false positives occurring adjacent to noise windows due to spreading of the impulsive noise during the decomposition.

Conclusions:
Effective removal of artifacts from CVP waveforms enables the extraction of meaningful features in the signals. The algorithm developed is patient-specific and fast enough for online event detection, allowing this general strategy to apply to any CVP signal regardless of the underlying cardiac anomalies.