BACKGROUND:Early recognition and intervention of hemorrhage are associated with decreased morbidity in children. Triage models have been developed to aid in the recognition of hemorrhagic shock after injury but require complete data and have limited accuracy. To address these limitations, we developed a Bayesian belief network, a machine learning model that represents the joint probability distribution for a set of observed or unobserved independent variables, to predict blood transfusion after injury in children and adolescents.
METHODS:We abstracted patient, injury, and resuscitation characteristics of injured children and adolescents (age 1 to 18 years) from the 2017 to 2019 Trauma Quality Improvement Project database. We trained a Bayesian belief network to predict blood transfusion within 4 hours after arrival to the hospital following injury using data from 2017 and recalibrated the model using data from 2018. We validated our model on a subset of patients from the 2019 Trauma Quality Improvement Project. We evaluated model performance using the area under the receiver operating characteristic curve and calibration curves and compared performance with pediatric age-adjusted shock index (SIPA) and reverse shock index with Glasgow Coma Scale (rSIG) using sensitivity, specificity, accuracy, and Matthew's correlation coefficient (MCC).
RESULTS:The final model included 14 predictor variables and had excellent discrimination and calibration. The model achieved an area under the receiver operating characteristic curve of 0.92 using emergency department data. When used as a binary predictor at an optimal threshold probability, the model had similar sensitivity, specificity, accuracy, and MCC compared with SIPA when only age, systolic blood pressure, and heart rate were observed. With the addition of the Glasgow Coma Scale score, the model has a higher accuracy and MCC than SIPA and rSIG.
CONCLUSION:A Bayesian belief network predicted blood transfusion after injury in children and adolescents better than SIPA and rSIG. This probabilistic model may allow clinicians to stratify hemorrhagic control interventions based upon risk.