Patent ductus arteriosus (PDA) is a condition in which the ductus arteriosus fails to close after birth. Hemodynamically significant patent ductus arteriosus (hsPDA) is a common complication in preterm infants, and its incidence is associated mainly with gestational age and birthweight (1). Persistent hsPDA (through which a continuous large right-to-left shunt occurs) can cause congestion of the pulmonary circulation and ischemia in the systemic circulation, thereby leading to dysfunction of multiple organs and increasing the risk of death (2).

Nonsteroidal anti-inflammatory drugs (NSAIDs), particularly ibuprofen and indomethacin, are first-line treatment to close hsPDA for preterm infants (3). The main mechanism of action of NSAIDs is to inhibit the active site of cyclooxygenase to reduce prostaglandin synthesis, thereby promoting constriction or closure of the ductus arteriosus (4). However, the ductal smooth muscle is immature in preterm infants. Hence, they may respond poorly to drugs, and have a risk of failure of PDA closure that requires a second type of treatment (medication or surgery). Various factors have been shown to influence the efficacy of NSAIDs for closing PDA, such as maternal-health status (5), gestational age (6), age in days (7), platelet count (8), and the type and dose of drugs (9). To reveal variables with nonlinear and complex relationships with the treatment outcome of NSAIDs in premature infants with hsPDA, an effective method to develop accurate predictive models is needed.

“Machine learning” is an important branch of artificial intelligence. It has been used widely in biomedicine and medicine, such as clinical diagnosis, precise treatment, and health monitoring. Machine learning involves learning and utilizing patterns in given data using advanced algorithms for making decisions or predictions about real-world events (10). However, models of machine learning are most often “black boxes” that cannot be directly interpretable to humans. Lack of interpretability undermines clinicians' trust in these black-box models, but also limits their operability for clinical prediction (11). Interpretable machine learning can provide understandable explanations for how a model works and why it makes specific predictions. It can help “debugging” of a model, direct collection of future data, provide reliable information for feature construction and human decision-making and, eventually, build trust between humans and models (12).

Here, we assessed 182 infants receiving NSAIDs for the first time to close PDA to develop an interpretable model of machine learning. This model aimed to predict the efficacy of NSAIDs for treating hsPDA in premature infants, and explain the relationship between clinical features and the treatment outcome. We hope this model can provide guidance for the clinical treatment of premature infants with PDA.

With the approval of the Institutional Review Board (IRB No. 2022-IRB-194), we evaluated preterm infants (gestational age ≤30 weeks) who were first treated with NSAIDs to close PDA at two tertiary medical centers in different regions of Zhejiang Province, China. Patients were excluded if they had one of the following conditions: (i) spontaneous closure of PDA after hospital admission; (ii) drug contraindications that necessitated direct surgical ligation; (iii) a history of NSAIDs treatment in other hospitals; (iv) complex congenital heart disease except PDA, atrial septal defect or small ventricular septal defect; and (v) incomplete data. All infants were treated with ibuprofen suspension or indomethacin.

Patients treated at ZCH between August 2015 and July 2022 were set as the “development cohort”. Two patients treated at Yiwu Branch Of Children's Hospital Zhejiang University School Of Medicine (YBZCH) in July 2022 were set as the “validation cohort”. The graphic abstract is shown as Figure 1.

In this retrospective cohort study, we developed and validated a machine-learning algorithm that used 11 maternal features, 20 neonatal features, and 8 laboratory and echocardiographic features to predict the efficacy of first-time treatment with NSAIDs to close hsPDA in preterm infants. The random forest model demonstrated good predictive performance and clinical interpretability. Among the top-10 features in the plot for variable importance, over half were laboratory and echocardiographic features. We illustrated how the top-three important features affected the treatment outcome using a marginal effect plot. Furthermore, we validated this predictive model with two external cases from another medical center. All the features used in the model are routine clinical variables collected before medication use. Therefore, the model can be helpful in clinical decision-making for preterm infants in the neonatal intensive care unit, which demonstrates the potential clinical value of the model.

NSAIDs have been used widely as alternatives to surgical PDA ligation for preterm infants. However, 10%–40% of patients obtain no benefits from such drugs (5). Some cohort studies or studies using regression analysis have discovered various maternal and neonatal clinical features that could be used to predict the likelihood of PDA closure with NSAIDs, such as gestational hypertension, gestational age, twin pregnancy, respiratory distress syndrome, age in days, and prenatal exposure to drugs such as magnesium sulfate and NSAIDs (5, 7, 14–17). Changes in urine volume or pulse pressure before and after NSAIDs treatment are not useful predictors of the responsiveness of preterm infants with PDA to a drug (18, 19). Those findings provide a reference for early recognition of the outcome of medication therapy and subsequent individualized management strategies for PDA. In addition, bedside ultrasonography can be used to assess the hemodynamic changes associated with the development, progression, and treatment of hsPDA (20, 21). Accordingly, we developed a machine-learning system to predict the probability of successful closure of PDA based on clinical features as well as laboratory and echocardiographic parameters 72 h before treatment. Our system demonstrated a reliability to predict the efficacy of NSAIDs for PDA closure in preterm infants of gestational age ≤30 weeks.

We wished to minimize the black-box effect and increase the transparency of our model. Hence, we interpreted the variable importance of the predictive model using plots for variable importance and marginal effect. According to these plots, the plasma albumin level and platelet count within 72 h before treatment and 24-h urine volume before treatment affected the therapeutic effect significantly. These exploratory findings may improve understanding of the features linked with the success or failure of NSAIDs therapy for preterm infants with PDA. Our findings may be useful for formulating strategies to reduce the risk of treatment failure in the future.

The relationship between the plasma albumin level and efficacy of NSAIDs for PDA closure has been reported rarely. NSAIDs have high affinity with plasma proteins (especially albumin) and bind to them in a reversible manner, which can affect the dose–response relationship and the therapeutic effect of drug combinations (22). With regard to hypoproteinemia, the free drug concentration in plasma increases, which can enhance drug efficacy and even cause adverse reactions but, simultaneously, alter the half-life and duration of action of the drug. We found that the efficacy of NSAIDs for PDA closure improved with an increase in the plasma albumin level within a certain range, further indicating that the efficacy of the drug may be dependent upon its duration of action.

Studies have demonstrated the association between the platelet count and response to a drug in treatment of hsPDA. A low platelet count is associated with a higher possibility of failure of PDA treatment using ibuprofen or indomethacin, whereas a high platelet count (≥181 × 10⁹/L) independently increases the probability of success of treatment with ibuprofen (8, 23). However, a recent study showed that increasing the platelet count to relatively high levels by transfusion to correct thrombocytopenia did not aid PDA closure, and instead increased the risk of intraventricular hemorrhage (24). Those contradictory results could be explained by our marginal-effect plot of model variables. The probability of success with NSAIDs was low when the platelet count <100 × 10⁹/L, and increased correspondingly if the platelet count was within 100–175 × 10⁹/L; but if platelet count continued to increase, the probability of treatment success did not increase further. Further research is needed to elucidate the mechanism of PDA closure and explore the potential value of platelet transfusion for improving PDA closure using medications.

Indomethacin and ibuprofen are nonselective inhibitors of cyclooxygenase. They can cause renal vasoconstriction, reduced renal blood flow, and acute kidney injury. NSAIDs-associated acute kidney injury can manifest as reduced urine volume and/or increased serum levels of urea/creatinine (25, 26). Therefore, reduced urine volume is considered to be a potential predictor of the efficacy of NSAIDs for closing PDA. We revealed that the 24-h urine volume before treatment was an important variable affecting the treatment outcome of NSAIDs. The probability of treatment success was low if the urine volume was 3.5–5.5 ml/kg·h. One study reporting on the relationship between urine-volume changes and PDA closure after indomethacin treatment in preterm infants concluded that urine-volume changes could not be used to predict therapeutic effects. Therefore, whether changes in urine volume can predict the therapeutic effects of NSAIDs merits further investigation (18).

Our study had several limitations. First, the model is restrospective and focused only on preterm infants who received NSAIDs treatment regimens for the first time to close hsPDA, not including with those in the literature described opportunities of escalation i.e., like increasing the dose of ibuprufene or shortening of the time intervall for indomethacine. Second, plasma albumine is not very common as a parameter in laboratory diagnostics today, especially in times, where less blood samples are made in premature infants. Finally, the sample size for external validation was extremely small due to the small number of premature infants admitted in another center and the incomplete data, which may have influenced the performance of the predictive machine-learning model. Large-sample prospective studies are needed to investigate the application of interpretable machine learning-based predictive models for improving the treatment outcome of preterm infants with hsPDA in clinical practice.

Based 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.