651 - Predictive model for rebound hyperbilirubinemia in babies born less than 35 weeks' gestation: A feasibility study

Publication Number: 1628.651

Mehwish Sheikh, University of illinois, Peoria, IL, United States; Shoji O. Samson, University of Illinois College of Medicine, Peoria, Peoria, IL, United States; Collins Odhiambo, UI Health, Peoria, IL, United States; Adam Cross, University of Illinois College of Medicine, Peoria, IL, United States

Background: Rebound hyperbilirubinemia (RH) is defined as a total serum bilirubin (TsB) level that reaches the phototherapy threshold within 72-96 hours of discontinuing phototherapy. The incidence of RH varies widely from 13.3% to 24.9%. Preterm population is inherently at greater risk of developing bilirubin-induced neurologic dysfunction (BIND) at lower TsB levels due to deficient mechanisms for bilirubin metabolism. A machine-learning (ML) model specifically designed for the preterm population may offer standardized guidance in monitoring rebound hyperbilirubinemia. Such information may not only enhance our understanding of hyperbilirubinemia in this population but also facilitate the integration of machine learning into the NICU workflow.

Objective: We aim to develop a prediction model for post-phototherapy serum bilirubin levels in babies < 35 weeks' gestation.

Design/Methods: After IRB approval, data were collected on neonates < 35 weeks of gestation from a single center-level IV unit between 01/2010 and 01/2022. Neonates with at least two bilirubin levels in the 72 hours of discontinuation of phototherapy were included. Neonates with significant congenital anomaly, intestinal pathology in the first 10 days, or with a single TsB in 72 hours post-phototherapy were excluded. Data from 01/2021-01/2022 was set aside for a hold-out set to analyze model performance.

Results: The original dataset included 3,663 neonates. After selecting cases with complete phototherapy data, the dataset was reduced to 1,335 records. Key descriptive statistics demonstrated a mean birth weight of 1,559 grams and a mean gestational age of 30.8 weeks. The mean duration between discontinuation of one phototherapy session and initiation of the next was approximately 60.4 hours (median 60.5, SD 2.95). The average time for bilirubin tests post-phototherapy was 7.4 hours. Eight machine learning models were assessed across four longitudinal time points. Among these, Random Forest achieved the highest average accuracy (0.881) and F1 score (0.89). Using SHAP analysis with the Random Forest fallback model, pre-phototherapy TsB emerged as the most influential predictor of hyperbilirubinemia rebound, standing out by a wide margin compared to other variables. Several weight-related features were also among the top contributors. Race, Parity, and Mode of Delivery show a modest yet meaningful influence on predicted rebound risk.

Conclusion(s): Pre-phototherapy total serum bilirubin is the most important predictor of RH in preterm neonates less than 35 weeks, followed by daily weight changes.

Hyperbilirubinemia trend across time points.
The prevalence of hyperbilirubinemia steadily decreases across the four time points. This pattern indicates a progressive decline in hyperbilirubinemia, with the most pronounced reduction occurring between the third and fourth measurements.

Comparison of Predictive models
Eight machine learning models were assessed for predicting rebound hyperbilirubinemia across four longitudinal time points, including Logistic Regression, Random Forest, Gradient Boosting, HistGradientBoosting, Extra Trees, K-Neighbors, Support Vector Classifier, and a baseline Dummy model. The models offered a strong balance between accuracy and subset accuracy. Overall, Random Forest emerged as the top-performing model based on accuracy, but ensemble methods such as Gradient Boosting and HistGradientBoosting also showed robust performance across multiple evaluation metrics, making them viable alternatives for prediction.

SHAP analysis