Background: Adults presenting with a neutrophil-predominant leukocytosis (white cell count >50,000/μL) often necessitate urgent medical management. These patients are diagnosed with either acute presentations of chronic myeloid malignancies or leukemoid reactions, yet accurate models to distinguish between these entities do not exist. We used demographic and lab data to build a machine learning model capable of discriminating between these diagnoses.
Methods: The medical record at a tertiary care medical center was queried to identify adults with instances of white counts greater than 50,000/μL and >50% neutrophils from 2000 to 2021. For each patient, a full set of demographic and lab values were extracted at the time of their first presentation with a white count >50,000/μL. We generated a series of models in which the parameters most predictive of myeloid malignancies were identified, and a supervised machine learning approach was applied to the dataset.
Results: Our best model-using a support vector machine algorithm-produced a sensitivity of 96% and a specificity of 95.9% (area under the curve = 0.982) for identifying myeloid malignancies. We also identified a clinically meaningful and significant disparity in outcomes based on diagnosis-a 6-fold increase in 12-month mortality in those diagnosed with leukemoid reactions.
Conclusions: These findings need to be validated but fill an unmet need for timely and accurate diagnosis in the setting of profound, neutrophil-predominant leukocytosis and support the use of predictive models as a means to improve patient outcomes.
Keywords: Leukemoid reaction; Machine learning; Myeloid malignancy.
Copyright © 2024 Elsevier Inc. All rights reserved.