Catheter related thrombosis in hospitalized infants: A neural network approach to predict risk factors

Abstract

IntroductionWe sought to investigate the predictors of catheter-related thrombosis (CRT) in a cohort of critically ill hospitalized infants and using a novel approach (the artificial neural network - ANN) in combination with conventional statistics to identify/confirm those predictors. MethodsWe performed a retrospective analysis of all infants with a central or peripherally inserted central venous catheter (CVC/PICC) between 2015 and 2018. ANN was generated to investigate the predictors of CRT. The predictive variables examined in the ANN were age, gender, weight, co-morbid conditions, line type, use of ultrasound (USG), emergent line placement, location of line tip, any major surgical procedures, use of mechanical ventilation, exposure to cardio-pulmonary bypass (CPB), past-history of CVC/PICC, or thrombosis. Binary logistic regression was performed to calculate odds ratios (ORs) and determine which factors were significant in predicting CRT. ResultsOf total of 613 infants, 59.9% of patients had a history of previous CVC or PICC and 12.2% had a history of thrombus as documented by USG in the past three months. CPB exposure was present in 48.1%. The incidence of CRT was 10.7%. Independent predictors of CRT were the line tip in IVC (OR: 2.37, 1.08–5.21, P = 0.032), history of thrombosis (OR: 2.40, 1.16–4.96, P = 0.019), previous CVC/PICC (OR: 2.80, 1.24–6.33, P = 0.014) and exposure to CPB (OR: 2.749, 1.08–6.98, P = 0.034). A sensitivity analysis was performed to determine the normalized importance of each variable used to create the ANN. The most important variables were age (with normalized importance of 100%), history of thrombosis, weight, and exposure to CPB (normalized importance of 68.2%). ConclusionsNearly 1 in 10 infants developed CRT. We found that catheter tip in IVC, exposure to CPB, history of vein thrombosis and history of CVC/PICC placement in the past 3 months are independently associated with a higher risk of CRT in infants by using conventional and neural network methods.