Construction and temporal external validation of interpretable machine-learning models for predicting tigecycline-associated hypofibrinogenemia.

background and purposeThere is a paucity of available clinical tools with which to accurately predict the risk of tigecycline-associated hypofibrinogenemia, an adverse reaction with a high incidence and serious consequences. This study aimed to developed an optimal machine-learning model for predicting tigecycline-associated hypofibrinogenemia and to facilitate its clinical application.

methodsA total of 896 inpatients who received tigecycline treatment in a large tertiary teaching hospital between 2016 and 2022 were included in the model development cohort. Subsequent temporal external validation of the models was performed on 313 patients admitted in 2023. Three robust machine-learning algorithms, including LASSO, Boruta, and VSURF were integrated to identify predictors. These features were then utilized to construct nine machine-learning models. The optimal one was identified through model evaluation and validation and was interpreted based on the SHAP method.

resultsThe three algorithms identified five predictors, including age, treatment duration, pre-dose fibrinogen level, D-dimer level, and activated partial thromboplastin time. The XGboost model was identified as the most suitable due to its stability and excellent discrimination, calibration, and clinical utility. The area under the receiver operating characteristic curve for the model were 0.823 (0.794-0.853), 0.810 (0.746-0.873), and 0.773 (0.721-0.825) in the training, internal validation, and temporal external validation cohorts, respectively.

conclusionsMachine-learning models are promising in solving the classification prediction problem of tigecycline-associated hypofibrinogenemia. Model interpretation based on SHAP enables clinicians to visualize individualized risk and tailor prophylactic strategies, potentially improving patient safety and treatment outcomes.