Shall we repeat? Predicting the necessity to repeat blood testing in pediatric emergency departments of laboratory tests previously performed in the community using a machine learning model

Community‑ordered laboratory blood tests are frequently re-ordered when the patients arrive at the pediatric emergency departments. When only a few hours separate the initial community blood tests from assessment in the emergency department, repeated laboratory testing often does not contribute to clinical management. Additional testing can inconvenience patients, increase out-of-pocket costs for patients and families, add costs to the healthcare system, and delay clinical decision-making. To address this issue, we collected community-issued laboratory tests for complete blood count, electrolytes, and C-reactive protein for children aged three months to 18 years treated in pediatric emergency departments within a nationwide Health Maintenance Organization. These were compared to repeated laboratory tests performed in the pediatric emergency department within 12 hours. Using these data, we first analyzed the frequency of the justified repeated testing according to pre-set criteria. Next, we developed a decision-support machine learning model to predict whether a laboratory test should be repeated, leveraging the patient’s prior community laboratory results and sociodemographic features. For complete blood count, electrolytes, and C-reactive protein, we found that only 13.3%, 16.3%, and 19.0% of the repeated tests, respectively, were deemed justified. Moreover, we show that an out-of-the-box XGboost model can predict the necessity for repeating these laboratory tests with 91.8%, 89.1%, and 79.2% accuracy, respectively. In the absence of false negatives, a condition of high clinical importance, the model maintains accuracies of 73.9%, 67.3%, and 55.4% for the three laboratory categories, respectively. This performance level remains sufficient to substantially reduce unnecessary repeat testing for complete blood count and electrolyte tests, but less so for C-reactive protein. The benefits of the model in clinical practice should be further studied.