Application of Machine Learning Models to Predict Tigecycline Exposure in Critically Ill Patients Using Different Limited Sampling Strategies.

Abstract

Background: The high variability in the pharmacokinetics of tigecycline may impact therapeutic efficacy, particularly in critically ill patients. This study introduces machine learning (ML) methods, aiming to clarify their effectiveness in individualized drug exposure prediction and determine the optimal modeling strategy.

Methods: The least absolute shrinkage and selection operator (LASSO) method was used to screen the covariates. Four ML models, random forest, extreme gradient boosting, multilayer perceptron, and long short-term memory network (LSTM), were trained and tested using 751 concentrations from 98 patients in the pharmacokinetic dataset to predict the area under the steady-state plasma concentration-time curve within 24 hours (AUC_0-24h) of tigecycline. Furthermore, on the internal test set, we evaluated the accuracy of the published tigecycline Population Pharmacokinetics (PopPK) models.

Results: Seven covariates were screened out by LASSO method, including total bilirubin, creatinine clearance, Sequential Organ Failure Assessment (SOFA) score, body mass index, serum creatinine, age, and prothrombin time. Internal validation showed that, among the four ML models, the LSTM model performed the best, with the combination of covariates and the plasma concentration feature "C₀ + C_0.5 + C₆" exhibited the optimal predictive performance (R²: 0.887, RMSE: 2.496 mg·h/L, F₂₀: 75%). In comparison, the model combining covariates with C₀ was the most concise and better suited to practical computational conditions; its LSTM model was comparable to some PopPK models in terms of overall trend fitting and systematic bias control (R²: 0.818, RMSE: 3.164 mg·h/L, F₂₀ = 60%).

Conclusion: The LSTM model established in this study enables an accurate estimate of tigecycline AUC_0-24h. Three-point sampling provided the lowest prediction error, whereas C₀ alone reduced the sampling burden. These findings provide an effective modeling reference for tigecycline therapeutic drug monitoring.