Frontiers in Pediatrics (Apr 2023)
An interpretable machine-learning model for predicting the efficacy of nonsteroidal anti-inflammatory drugs for closing hemodynamically significant patent ductus arteriosus in preterm infants
Abstract
BackgroundNonsteroidal anti-inflammatory drugs (NSAIDs) have been widely used in the closure of ductus arteriosus in premature infants. We aimed to develop and validate an interpretable machine-learning model for predicting the efficacy of NSAIDs for closing hemodynamically significant patent ductus arteriosus (hsPDA) in preterm infants.MethodsWe assessed 182 preterm infants ≤ 30 weeks of gestational age first treated with NSAIDs to close hsPDA. According to the treatment outcome, patients were divided into a “success” group and “failure” group. Variables for analysis were demographic features, clinical features, as well as laboratory and echocardiographic parameters within 72 h before medication use. We developed the machine-learning model using random forests. Model performance was assessed by the area under the receiver operating characteristic curve (AUC). Variable-importance and marginal-effect plots were constructed to explain the predictive model. The model was validated using an external cohort of two preterm infants who received ibuprofen (p.o.) to treat hsPDA.ResultsEighty-three cases (45.6%) were in the success group and 99 (54.4%) in the failure group. Infants in the success group were associated with maternal chorioamnionitis (p = 0.002), multiple births (p = 0.007), gestational age at birth (p = 0.020), use of indometacin (p = 0.007), use of inotropic agents (p < 0.001), noninvasive ventilation (p = 0.001), plasma albumin level (p < 0.001), PDA size (p = 0.038) and Vmax (p = 0.013). Multivariable binary logistic regression analysis showed that maternal chorioamnionitis, multiple births, use of indomethacin, use of inotropic agents, plasma albumin level, and PDA size were independent risk factors influencing the efficacy of NSAIDs (p < 0.05). The AUC of the random forest model was 0.792. The top-three features contributing most to the model in the variable-importance plot were the plasma albumin level and platelet count 72 h before treatment and 24-h urine volume before treatment. In the external cohort, treatment succeeded in one case and failed in the other. The probabilities of success and failure predicted by the random forest model were 60.2% and 48.4%, respectively.ConclusionBased on clinical, laboratory, and echocardiographic features before first-time NSAIDs treatment, we constructed an interpretable machine-learning model, which has a certain reference value for predicting the closure of hsPDA in premature infants under 30 weeks of gestational age.
Keywords
