Machine Learning to Improve Accuracy of Transcutaneous Bilirubinometry.

Abstract

Introduction: This study aimed to develop models for predicting total serum bilirubin by correcting errors of transcutaneous bilirubin using machine learning based on neonatal biomarkers that could affect spectrophotometric measurements of tissue bilirubin.

Methods: This retrospective study included infants born at our hospital (≥36 weeks old, ≥2,000 g) between January 2020 and December 2022. Infants without a phototherapy history were included. Robust linear regression, gradient boosting tree, and neural networks were used for machine learning models. A neural network, inspired by the structure of the human brain, was designed comprising three layers: input, intermediate, and output.

Results: Totally, 683 infants were included. The mean (minimum-maximum) gestational age, birth weight, participant age, total serum bilirubin, and transcutaneous bilirubin were 39.0 (36.0-42.0) weeks, 3,004 (2,004-4,484) g, 2.8 (1-6) days of age, 8.50 (2.67-18.12) mg/dL, and 7.8 (1.1-18.1) mg/dL, respectively. The neural network model had a root mean square error of 1.03 mg/dL and a mean absolute error of 0.80 mg/dL in cross-validation data. These values were 0.37 mg/dL and 0.28 mg/dL, smaller compared to transcutaneous bilirubin, respectively. The 95% limit of agreement between the neural network estimation and total serum bilirubin was -2.01 to 2.01 mg/dL. Unnecessary blood draws could be reduced by up to 78%.

Conclusion: Using machine learning with transcutaneous bilirubin, total serum bilirubin estimation error was reduced by 25%. This integration could increase accuracy, lessen infant discomfort, and simplify procedures, offering a smart alternative to blood draws by accurately estimating phototherapy thresholds.