352 - Non-invasive estimation of Total Serum Bilirubin from clinical parameters using Machine Learning

Poster Number: 352
Publication Number: 352.123

Juhee Lee, New York Medical College, NEW YORK, NY, United States; Crhistian Toribio Dionicio, New York Medical College, New York, NY, United States; Ruth Vela Sosa, New York Medical College, new york, NY, United States; Gayathri Sreenivasan, New York Medical College, New york, NY, United States; Trusha M. Patel, New York Medical College, New York, NY, United States; Arpit Gupta, New York Medical College, New York, NY, United States

Background: Neonatal hyperbilirubinemia is a common pathology of the newborn, affecting nearly half of all term infants, with serious complications such as neurotoxicity when not recognized and treated early.  
Conventionally, serum bilirubin level is measured by heel puncture - a procedure that is painful for the infant, and anxiety-producing for the family. Transcutaneous bilirubin (TcB) measurement is a well-established method to estimate bilirubin level non-invasively through the skin. However, TcB levels can have variations up 2 to 3 mg/dL from total serum bilirubin (TSB) levels. Due to the wide error range of TcB, bilirubinometers cannot totally substitute for blood sampling of the infants.
Machine Learning is a powerful set of tools that can identify complex patterns within heterogenous data. One prior study utilized machine learning to predict initiation of phototherapy with TSB values. However, to the best of our knowledge, there has been no literature for direct prediction of TSB using routinely collected data in newborns.

Objective: Our goal was to predict TSB values using TcB measurements and data collected from the Electronic Healthcare Record (EHR) using machine learning.

Design/Methods: This was a single-center retrospective study of neonates with gestational age more than 35 weeks who had concomitant measurements of TcB and TSB in the Nursery and Neonatal Intensive Care Unit between December 2019 and November 2021. We collected a total of 32 variables including the TcB values, which are listed in Table 1. Exclusion criteria were newborns who stayed longer than 5 days or who received phototherapy.
We developed an XGBoost regression model using TcB value, in addition to demographic and clinical parameters present in the EHR. We applied a 10-fold cross-validation strategy to ensure capture of variability within the dataset. Model performance was evaluated by calculation of the Mean Absolute Error (MAE).

Results: We obtained TcB, TSB and EHR measurements for 353 newborns. MAE for the XGBoost Regression model was 0.64 (95% CI: 0.56 – 0.72). In comparison, MAE for TcB was 2.0 (95% CI: 1.86 – 2.14) (Figure 1). Majority of individual TSB predictions were seen to lie on the line of perfect concordance across 10-cross-validation folds (Figure 2). Conclusion(s): Machine learning enabled more accurate estimation of TSB with easily available EHR data and TcB values, compared to traditional bilirubinometry alone. Machine learning can be used to augment non-invasive bilirubinometry and reduce reliance on invasive tests.
Table 1. Collected variables in the dataset.Categories and the variables that are included in the dataset.
Figure 1. A density graph for absolute differenceComparison of MAE distribution between TcB and TSB, and an XGBoost regression model for TSB estimation.