RWE-007 - PREDICTING LINEZOLID-INDUCED HEMATOLOGIC TOXICITIES IN REAL-WORLD PATIENTS WITH HYBRID MODELING.

A. Patel¹, S. Sun¹, A. Butte², K. Radtke¹; ¹University of California, San Francisco, San Francisco, CA, USA, ²Bakar Computational Health Sciences Institute, San Francisco, CA, USA.

Background: Linezolid (LZD) is a crucial antibiotic in fighting rising antimicrobial resistance, but serious hematologic toxicities limit its use. Knowledge is still lacking on its optimal use to avoid safety events in real-world settings and when pharmacokinetic data is not available. Here, we aim to use machine learning and survival analysis to predict patients at risk of adverse events (AEs) with real-world data.

Methods: De-identified electronic health records (EHRs) were used to identify patients treated with LZD. The risk period was from LZD start date to 14 days after last dose. Hematologic AEs encompassed grade 3+ thrombocytopenia, grade 3+ anemia, platelet transfusion, and red blood cell transfusion. Demographic, drug exposure, laboratory, and other clinical features were extracted from EHRs. Random forest classifier (RFC) with 80/20 train/test split was used to predict any hematologic AE and to identify important features for multivariate Cox survival analysis.

Results: Of 1995 patients included, 817 (45%) had at least one AE. RFC predictive accuracy was 79% (auROC 0.80) with all 41 collected features and 77% (auROC 0.77) with the 5 most important features: baseline platelet (Plt), aspartate aminotransferase (AST), serum creatinine (Scr), hemoglobin, and age. From Cox analysis with these 5 features, significant risk factors were baseline Plt count < 150 x10^9/L (hazard ratio (HR): 6.45, 95% CI: 5.45-7.62), AST > 5x upper limit of normal (HR: 1.60, 95% CI: 1.29-1.97), and Scr 3+ mg/dL (HR: 1.63, 95% CI: 1.36-1.96).

Conclusion: Patients at risk of hematologic AEs with LZD can be accurately predicted with 5 factors that are routinely collected in clinical settings. This work shows potential for EHR data-driven clinical decision support tools for the safe prescribing and management of drugs with known and measurable toxicities.