Bloodstream Infections in Childhood Acute Myeloid Leukemia and Machine Learning Models: A Single-institutional Analysis

Abstract

Childhood acute myeloid leukemia (AML) requires intensive chemotherapy, which may result in life-threatening bloodstream infections (BSIs). This study evaluated whether machine learning (ML) could predict BSI using electronic medical records. All children treated for AML at Children’s Minnesota between 2005 and 2019 were included. Patients with Down syndrome AML or acute promyelocytic leukemia were excluded. Standard statistics analyzed predictors of BSI, and ML models were trained to predict BSI. Of 95 AML patients, 54.7% had BSI. Of 480 admissions, 19% included BSI. No deaths were related to BSI, and survival of non-Whites was significantly inferior to White patients. Logistic regression revealed that higher cytarabine doses increased the risk of BSI, with an odds ratio (OR) of 1.110 (P < 0.05). Prophylactic levofloxacin-vancomycin reduced the risk of BSI, with OR of 0.495 (P < 0.05). The best-performing ML model was regularized logistic regression with an area under the curve (AUC) of 0.748, improved specificity by 37.5% compared with neutropenia, and 2.6% compared with fever. In conclusion, BSI risk was increased by cytarabine and reduced by levofloxacin-vancomycin prophylaxis. ML predicted BSI with improvement over fever or neutropenia. In clinical practice, ML may offer flexibility by controlling sensitivity and specificity by adjusting BSI diagnosis thresholds.