Prediction of methotrexate neurotoxicity using clinical, sociodemographic, and area-based information in children with acute lymphoblastic leukemia.

Abstract

Background: Methotrexate is a critical component of pediatric acute lymphoblastic leukemia (ALL) therapy that can result in neurotoxicity which has been associated with an increased risk of relapse. We leveraged machine learning to develop a neurotoxicity risk prediction model in a diverse cohort of children with ALL.

Methods: We included children (age 2-20 years) diagnosed with ALL (2005-2019) and treated in Texas without pre-existing neurologic disease. Clinical information was obtained by medical record review. Neurotoxicity occurring post-induction and prior to maintenance therapy was defined as neurologic episodes occurring within 21 days of methotrexate. Suspected cases were independently confirmed by 2 pediatric oncologists. Demographic and clinical factors were compared using logistic regression. The dataset was randomly split (80/20) for training and testing. random forest (RF) with boosting and downsampling using 5-repeat, 10-fold cross-validation was used to construct a predictive model.

Results: Neurotoxicity developed in 115 (8.7%) of 1325 eligible patients. Several factors including older age at diagnosis (OR = 1.19, 95% CI: 1.15-1.24) and Latino ethnicity (OR = 2.79, 95% CI: 1.83-4.35) were associated with neurotoxicity. The RF had an area under the curve of 0.77 with a train error rate of 0.29 and a test error rate of 0.24. The overall sensitivity was 0.73, and specificity was 0.69.

Conclusions: In one of the largest studies of its kind, we developed a novel risk prediction model of methotrexate-related neurotoxicity. Ultimately, a validated model may help guide the development of personalized treatment strategies to reduce the burden of neurotoxicity in children diagnosed with ALL.