Patent ductus arteriosus (PDA) is a common problem in premature infants and is associated with mortality and neonatal morbidity. Approach to treatment selection (who, when and how) is controversial and polarizing opinion among neonatologists. Recent trials of early ibuprofen treatment have not shown efficacy in reducing the composite outcome of death and/or bronchopulmonary dysplasia.1, 2 Patient selection reflected real-world practice, selecting the eligible patient population based on PDA diameter > 1.5 mm and predominant left-right transcanal shunt, so treatment effects were not uniform. The question of whether the enrolled population reflects patients with moderate to high volume shunts is also questionable. The importance of patient selection was demonstrated in a single institution where low-volume PDA shunts (some of which were >1.5 mm in diameter) left untreated for a minimum of 2 weeks did not increase the risk of unfavorable outcomes.3 Additionally, targeted PDA therapy resulted in improvement of death, severe intraventricular hemorrhage, and grade 3 bronchopulmonary dysplasia when the hemodynamic significance was clearly defined.4 Controversies over definitions, approaches to screening, and management strategies strongly suggest the potential of artificial intelligence and machine learning models in identifying and predicting PDA.
In this issue, pediatric researchKun et al. 5 describe a machine learning-based tool that facilitates the recognition of premature infants at high risk of hemodynamically significant PDA (hsPDA). The final model included six variables: birth weight, mean fraction of inspired oxygen (FiO).2) during the first 3 days, mean blood pressure (BP) during the first 3 days, percentage of time saturation exceeded 96% during the first 3 days, amount of surfactant taken during the first 3 days, and method of administration. This model achieved sensitivity of 74.8% and specificity of 53.4% with an area under the ROC curve of 0.685. The construction of this model was based on the assumption that receiving treatment is a hallmark of hsPDA. This approach places great emphasis on the validity of the echocardiographic definition of hemodynamic significance. In routine clinical practice, determination of hemodynamic significance is based on the presence of a left-to-right PDA shunt and the identification of one of the following echocardiographic parameters: left pulmonary artery end-diastolic flow > 0.2 m/s, abdominal aorta/celiac trunk diastolic flow reversal, or left ventricular load (visual estimation of ejection fraction/left atrial-to-aortic ratio > 1.4/left ventricular end-diastolic diameter > 15 mm/kg). It has been previously reported that many echocardiographic measurements are prone to operator-dependent errors, impacting reliability.6 Misjudgment of hemodynamic significance can lead to both undertreatment and overtreatment. As an example, echocardiographic findings of left ventricular stress can also be seen in the setting of LV dysfunction and in patients with systemic hypertension, clinical scenarios in which PDA closure is not recommended. Drug therapy itself is not without side effects, and there is animal evidence of impaired angiogenesis.7 Therefore, it is of utmost importance to avoid unnecessary treatment in patients with low shunt volumes or those with a high probability of spontaneous closure. On the other hand, data from a preterm baboon model suggest that early ibuprofen improved alveolar formation.8 Therefore, a comprehensive hemodynamic evaluation is required to ensure that the echocardiographic markers are consistent with the findings of hsPDA and are not another disease process masquerading as hsPDA where only one or more markers are present.9 Screening high-risk patients using a multiparametric scoring system, such as the Iowa PDA Score (Figure 1), distinguishes between low-volume shunts that do not require treatment and intermediate- and high-volume shunts that require treatment, allowing for the presence of one or more markers without misdiagnosing hsPDA when it could be an entirely different diagnosis.10
Ao aortic root, LA left atrium, LVO left ventricular output, MCA middle cerebral artery, PDA patent ductus arteriosus, RVO right ventricular output.
Routine clinical evaluation has been shown to be unreliable in identifying hsPDA. Early clinical signs of large-volume shunts, such as increased oxygen and ventilation requirements, decreased diastolic and mean blood pressure, and metabolic acidosis, may be due to immaturity or other disease states. A prospective observational study of 154 patients showed that only metabolic acidosis consistently predicted the presence of major PDA at day 7.11. These data highlight current knowledge gaps and the potential benefits of using machine learning in hsPDA identification and prediction, a much-needed research area. Machine learning capabilities enable complex analysis that is not possible with commercially available monitoring techniques and current evaluation strategies. The authors chose to include only two continuously monitored variables in the final model. These were the mean blood pressure over the first 3 days and the percentage of time oxygen saturation was above 96% during the first 3 days. Although these data are valuable, using only averages of mean blood pressure over such long periods can also result in loss of signal that can reveal details in more complex analyses. In particular, differences in blood pressure measurement components (systolic and diastolic blood pressure) and location (preluminal and postluminal blood pressure).12 It may be possible to better distinguish physiological differences between groups. Additionally, dividing the evaluation period into smaller chunks for analysis may have provided signals useful for characterizing transluminal flow patterns and shunt volume. Unfortunately, the authors did not have continuous vital sign collection for analysis. However, they report that they get a value every 15 minutes. Another potential improvement in model performance could have resulted from arterial waveform analysis of the model with continuous blood pressure monitoring.
One of the challenges in creating and implementing machine learning models like the one described in the paper is that they rely on worsening clinical symptoms to identify hsPDA. Echocardiographic findings identifying hsPDA have been shown to precede clinical signs by up to 4 days.13 Wait until infant requires high FiO2 Considering the use of this tool to evaluate hsPDA may delay diagnosis and increase the risk of undesirable outcomes. Additionally, infants with increased FiO2 It is possible that these two variables are collinear, as those who meet the requirements are more likely to receive higher doses of surfactant therapy. Furthermore, the clinical assumption that respiratory decline is exclusively pulmonary in origin may lead to delayed diagnosis and the application of biologically counterintuitive or harmful treatments. For example, in older preterm infants (>6 days old), PDA status is an important determinant of late surfactant response. Specifically, infants with a small or closed PDA are more likely to have a positive response, whereas infants with a PDA of 1 mm or more are more likely to have a negative response.14 In institutions without established screening programs for hemodynamic status, the likelihood of having hsPDA treated with surfactants (or other therapies) is significantly increased due to the lack of diagnostic accuracy for hypoxemia.
Although echocardiography improves diagnostic accuracy, it is not without limitations. Most hemodynamic centers are not routinely available 24/7, which can delay disease recognition and treatment. Second, the dynamic nature of cardiopulmonary physiology and the developmental vulnerability of immature infants are important considerations, as the underlying phenotype can change significantly and rapidly. For example, inhaled nitric oxide in preterm infants with pulmonary hypertension can produce significant changes in both pulmonary vascular resistance and the direction and magnitude of ductal shunts. Recognizing threshold points for changing phenotypes is difficult. Therefore, the timing of follow-up echocardiography is difficult. Machine learning facilitates enhanced monitoring and pattern recognition, which may lead to more accurate phenotypic characterization.
Overall, the authors are commendable for their investment in demystifying the diagnosis of hsPDA in this patient population using a relatively simple tool. However, it is important to recognize that hsPDA is a complex diagnosis, and simplifying the diagnosis can lead to both overdiagnosis and underdiagnosis, resulting in over- or under-treatment. Future research in the field of machine learning and predictive analytics should consider disease determination by ensuring the ability to distinguish between PDAs that are present but do not require treatment (low volume shunts) and PDAs that require treatment from patients without PDA (moderate or high volume shunts). Achieving this may require continuous vital sign monitoring, waveform analysis, and/or physiological variables such as echocardiographic markers of volume loading. The focus should continue to be on early identification of hsPDA and precision treatment to reduce adverse outcomes.
