Classification Tree Analysis Based On Machine Learning for Predicting Linezolid-Induced Thrombocytopenia

Abstract

Linezolid-induced thrombocytopenia is related to linezolid exposure, baseline platelet count and patient background. Although the relationship usually reflects the time of onset of thrombocytopenia, if the platelet maturation process is taken into account, the platelet decrease can be considered to have started at the beginning of treatment. To predict linezolid-induced thrombocytopenia, classification and regression tree (CART) analysis based on machine learning has been applied to identify predictive factors and cutoff values. We examined 74 patient data with or without linezolid-induced thrombocytopenia. Linezolid concentration and platelet count change, baseline platelet count, age, body weight and creatinine clearance estimate were evaluated as predictive factors for linezolid-induced thrombocytopenia. CART analysis selected the final tree containing two cutoff values: a platelet count reduction to less than 2.3% from baseline at 96 h after the initial dose and a linezolid concentration greater than or equal to 13.5 mg/L at 96 h after the initial dose. The targets for therapeutic intervention were concluded to be the linezolid concentration and the platelet change from baseline at 96 h after the initial dose. These cutoff values occur prior to the onset of thrombocytopenia and should be monitored to avoid linezolid-induced thrombocytopenia.

Keywords: Clinical pharmacokinetics; Machine learning; Pharmacodynamics; Pharmacokinetic/pharmacodynamic model; Pharmacokinetics; Population pharmacodynamics; Population pharmacokinetics.