Predicting tigecycline-related adverse events in infected patients: a machine learning approach with clinical interpretability

Abstract

Background: Tigecycline (TGC), while effective against multidrug-resistant infections, is limited by hepatotoxicity and coagulation disorders, yet lacks robust predictive tools. Methods: We developed an online dynamic nomogram to assess these adverse events using retrospective data from 2,553 TGC-treated patients (2020–2025). Seventy-seven clinical features were analyzed using Boruta and the Least Absolute Shrinkage and Selection Operator (LASSO) for feature selection. Seven machine learning (ML) models were evaluated via ten-fold cross-validation, as well as Receiver Operating Characteristic (ROC) curve and calibration curves, with SHapley Additive exPlanations (SHAP) analysis for interpretability and an online dynamic nomogram for clinical translation. Results: Logistic regression (LR) outperformed other algorithms, achieving Area Under the ROC Curve (AUC) values of 0.800 (95% CI: 0.727–0.874) for hepatotoxicity and 0.755 (95% CI: 0.665–0.845) for coagulation dysfunction. Independent risk factors for liver injury included prolonged treatment duration, high dosage, ICU admission, hepatitis B virus (HBV) infection, and elevated baseline levels of lactate dehydrogenase (LDH) and gamma-glutamyl transferase (GGT). Risk factors for coagulation dysfunction included extended treatment duration, ICU admission, elevated baseline creatinine (Cr), sepsis, and septic shock. Notably, co-administration of meloxicillin and higher baseline red blood cell (RBC) levels appeared to be protective. Conclusion: This study constructed an online dynamic nomogram with good discrimination and calibration, which can help to identify high-risk patients and assist clinicians in early risk stratification and individualized treatment planning.