191 - Development and Validation of a Novel Machine Learning Model to Predict Pharmacologic Closure of Patent Ductus Arteriosus

Publication Number: 191.4440

Puneet Sharma, Emory University School of Medicine, Atlanta, GA, United States; Addison Gearhart, Seattle childrens hospital, Renton, WA, United States; Guangze Luo, Harvard Medical School, Austin, TX, United States; Cindy Wang, McLean Hospital, Cambridge, MA, United States; Kristyn Beam, Beth Israel Deaconess Medical Center, Boston, MA, United States; Fotios Spyropoulos, Brigham and Women’s Hospital/Harvard Medical School, Boston, MA, United States; Andrew Powell, Boston Children's Hospital and Harvard Medical School, Boston, MA, United States; Philip Levy, Boston Children's Hospital, BROOKLINE, MA, United States; Andrew Beam, Harvard Medical School, Boston, MA, United States

Background: The patent ductus arteriosus (PDA) is associated with significant morbidity and mortality in the preterm infant. While pharmacotherapy can be effective, it is difficult to predict whether a patient will respond, leading to delays in care. Machine learning has emerged as a powerful tool to interpret clinical data to predict clinical outcomes but has not yet been applied to this question.

Objective: To train and validate a novel deep learning model to predict the likelihood of PDA closure after an initial course of pharmacotherapy in preterm infants.

Design/Methods: We identified a retrospective cohort of preterm infants ( < 37 weeks) admitted to the NICUs at Beth Israel Deaconess Medical Center (BIDMC) and Brigham and Women’s Hospital (BWH) between January 2016 and December 2021 who received pharmacologic treatment for their PDA. Infants were excluded if they received prophylactic indomethacin, had early termination of therapy, did not have an echocardiogram prior to therapy, or had other cardiac lesions. Relevant perinatal data and pre-treatment echocardiograms were collected. We randomized subjects into training and validation subgroups (70:30 split). We trained two distinct convolutional neural networks (CNN), one based on echocardiograms alone and the other on both echocardiograms and perinatal data. We compared the performance of the CNNs against controls of random forest and logistic regression models trained on perinatal data alone.

Results: We identified 174 infants who met our inclusion and exclusion criteria –104 from BIDMC and 70 from BWH. The success rate of pharmacotherapy was 60%. 121 infants were randomized to the training subgroup and 53 to the validation with no significant difference between the two. A total of 174 echocardiograms (1926 clips) were used. The training subgroup had 72% of clips and the validation had 28% of clips. The multimodal CNN had an AUC of 0.77, outperforming the imaging only model (AUC = 0.66). Additionally, the multimodal CNN outperformed logistic regression (AUC = 0.66) and random forest (AUC = 0.74) models.

Conclusion(s): Our novel, multimodal CNN shows promise for clinicians who do not currently have a reliable tool to predict success of pharmacologic closure. This investigation represents the first attempt to utilize machine learning methodology to predict this outcome.

Figure 1. Flowchart of patient enrollment and randomization
Figure 1 PDA.pdf

Table 1. Baseline characteristics of the randomized training and validation subgroups
Table 1 PDA.pdf