作者
Jieyu Sun,Yongkang Geng,Ting Zhou,Wei He,Ziyi Wang,Yuqiong Kuang,Yongqing Wang,Luning Sun,Mingqing Xu
摘要
Purpose: Linezolid exposure is highly variable in critically ill patients, increasing the risk of subtherapeutic or supratherapeutic concentrations. This study aimed to develop an interpretable machine learning model for classifying steady-state linezolid trough concentrations as subtherapeutic (<2 mg/L), target (2-8 mg/L), or supratherapeutic (>8 mg/L), and to assess model performance in a temporally separated test cohort. Patients and Methods: This single-center retrospective study included 299 adult intensive care unit patients who received linezolid and underwent therapeutic drug monitoring between June 2023 and March 2026. Missing data were handled using multiple imputation by chained equations, and Elastic Net was used for feature selection. Multinomial logistic regression, random forest, XGBoost, LightGBM, CatBoost, and multilayer perceptron models were compared using five-fold cross-validation across five imputed datasets. The best-performing model was interpreted using SHapley Additive exPlanations. The six models were additionally evaluated in a temporal test cohort. Results: Among the 299 patients, 37 (12.4%) had subtherapeutic, 186 (62.2%) target, and 76 (25.4%) supratherapeutic trough concentrations. Fifteen predictors were retained. CatBoost showed the best internal multiclass discrimination (AUROC 0.742, 95% CI 0.732-0.752; AUPRC 0.551, 95% CI 0.531-0.572). In the temporal test cohort, CatBoost achieved an AUROC of 0.754 (95% CI 0.684-0.826) and an AUPRC of 0.620 (95% CI 0.529-0.727). Estimated glomerular filtration rate was the most influential predictor, with additional contributions from inflammatory and hematologic indices, dosing variables, age, body mass index, and pathogen-related context. Conclusion: Routinely available clinical variables provided moderate discrimination among clinically distinct linezolid exposure categories, and performance was retained in a temporal test cohort. Because eGFR was the dominant predictor, the model may be most useful for flagging patients with altered renal function for earlier or more intensive TDM. Given its moderate discrimination and imperfect calibration, it should be regarded as a risk-stratification tool and should not be used for empirical dose modification without measured concentrations. Multicenter external and prospective validation is required before routine clinical implementation.