Machine Learning Classifier Using Blood Count Parameters and Erythropoietin to Predict JAK2 Mutations in Patients With Erythrocytosis.

Abstract

Differentiating polycythemia vera from other causes of erythrocytosis is a diagnostic challenge. Although most patients with polycythemia vera have Janus kinase 2 (JAK2) mutations, extensive testing is impractical because this is an uncommon cause of erythrocytosis. Identifying polycythemic patients most likely to benefit from JAK2 testing would improve use of this test. To develop an artificial intelligence analysis/machine learning classifier using blood count parameters and erythropoietin to predict JAK2 results in patients with erythrocytosis. Results from the Veterans Affairs data warehouse were used for training and validation. Cases with JAK2 results and hemoglobin values 15 g/dL or higher and 17 g/dL or higher in females and males, respectively, were included. Erythropoietin was optional. The highest performing model was evaluated with an out-of-sample data set. Among 31 models trained on data from 8479 individuals, including 540 (6.4%) positive for JAK2, Light Gradient Boosted Trees Classifier performed best. When applied to 330 out-of-sample cases with 9 (2.7%) positive for JAK2, the classifier's sensitivity, specificity, positive predictive value, and negative predictive value, were 100%, 92.8%, 28.1%, and 100%, respectively. Among a subset of 183 out-of-sample cases, the model's algorithm would have potentially reduced JAK2 testing by 89% compared with a 50% to 62% reduction using previously reported rule-based systems that similarly used blood count parameters. Platelet count had the greatest impact on the model, followed by relative distribution width and erythropoietin. These results show that a machine learning classifier may be beneficial as a decision support aid for JAK2 testing in polycythemic patients.