Exploring the Characteristics of Infants That Influence Their Number of Transfusions Using a Multivariable Multiclassification Model: A Retrospective Study

Abstract

Introduction: Factors that influence neonatal transfusions are poorly understood because of individual variations in birth conditions and maternal complications during pregnancy. This study aimed to establish models that incorporate perinatal factors associated with the early prediction and timely management of conditions of infants that require transfusion.

Methods: Data were collected from electronic medical records. Infants were categorized into non-transfusion, one transfusion, two transfusions, three transfusions, four transfusions, and more than four transfusions groups based on transfusions performed during hospitalization. Models were constructed to predict the number of transfusions needed by the infants using variables that showed significant differences among different transfusion groups based on multivariable, random forest, and gradient boosting tree multiclassification tasks.

Results: Underweight status, premature birth, Apgar scores at 1 min, 5 min, and 10 min, and gestational diabetes mellitus impacted the number of transfusions required by infants. The weighted macro-average area under the curve (AUC) values of three models constructed using previously mentioned variables were as follows: multivariable multiclassification model, AUC = 0.6549/0.7282/0.7379 on training/testing/validation sets; random forest multiclassification model, AUC = 0.8037/0.7628/0.7985 on training/testing/validation sets; and gradient boosting tree multiclassification model, AUC = 0.7422/0.7038/0.7488 on training/testing/validation sets. The weighted macro-average AUC of the three models constructed using Apgar scores were as follows: multivariable multiclassification model, AUC = 0.6425/0.7044/0.7379 on training/testing/validation sets; random forest multiclassification model, AUC = 0.7659/0.7662/0.7985 on training/testing/validation sets; and gradient boosting tree multiclassification model, AUC = 0.6559/0.6251/0.7488 on training/testing/validation sets.

Conclusion: Apgar scores at 1 min, 5 min, and 10 min may be preliminary predictive factors that could be used to implement early transfusion strategies for infants after birth. Because of the limitations of the data volume, variable selection, and model performance evaluation, further optimization and improvements are necessary to develop accurate blood transfusion prediction models for infants.

Keywords: Machine learning; Multiclassification model; Neonatal transfusion; Perinatal factor.