Prediction model of thrombocytopenia on patients using linezolid: a ranked-based covariate selection with...

Thrombocytopenia is a common adverse reaction of linezolid, often leading to severe complications. It is challenging to externally validate directly apply developed in other countries. We aim to develop and validate a risk prediction model of linezolid-associated thrombocytopenia (LAT) tailored to Vietnamese setting and to construct a simplified risk score calculation to support clinical decision-making.

Data was collected retrospectively from three large hospitals in Northern Vietnam. We selected inpatients treated with linezolid from November 2019 to March 2023. Potential predictors were chosen based on literature review and clinical experts' opinions. Final predictors were selected using Bayesian model selection. Thrombocytopenia was defined as platelet count value ≤112.5 G/L and a decrease more than 25% from the baseline. A multivariable logistic regression model was constructed to predict the occurrence of LAT. The final model was further validated using internal-external cross-validation.

Of 776 patients included, 247 patients (31.8%) developed LAT. Logistic regression model selection indicated that the risk predictors were age, duration of linezolid ≥14 days, baseline platelet count, creatinine clearance, sepsis, cirrhosis, continuous renal replacement therapy (CRRT) and heparin use. The model had moderate discrimination, with area under the curve (AUC) of 0.77 (95% confidence interval (CI): 0.72-0.83). Model calibration was good, with calibration-in-the-large and calibration slope of 0.00 (-0.38 to 0.38), and 0.92 (0.59-1.26) respectively. A risk score scale was established, with the optimal cut-off value being 23 points. Patients were categorised based on this score into three groups: low risk (-1 to 13 points), moderate risk (14-22 points) and high risk (≥23 points).

Our newly developed risk prediction model demonstrated moderate discriminatory ability in predicting the occurrence of LAT. From there, a simplified risk score was constructed to facilitate its applicability in clinical practice.