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

Area of Science:

• Pharmacovigilance and Machine Learning in Clinical Practice
• Predictive Modeling for Adverse Drug Reactions
• Antibiotic Safety and Risk Stratification

Background:

• Tigecycline-associated hypofibrinogenemia is a significant clinical concern with limited predictive tools.
• Accurate risk prediction is crucial for patient safety and optimizing tigecycline use.
• Developing clinical tools to forecast this adverse event is a pressing need.

Purpose of the Study:

• To develop and validate an optimal machine-learning model for predicting tigecycline-associated hypofibrinogenemia.
• To identify key predictors contributing to the risk of this adverse reaction.
• To facilitate the clinical application of a predictive tool for tigecycline therapy.

Main Methods:

• Utilized LASSO, Boruta, and VSURF algorithms to identify predictors from 896 patients (2016-2022) and validated on 313 patients (2023).
• Constructed nine machine-learning models, selecting the optimal one based on evaluation and validation.
• Employed SHAP for model interpretation to understand individual risk factors.

Main Results:

• Identified five key predictors: age, treatment duration, pre-dose fibrinogen, D-dimer, and activated partial thromboplastin time.
• The XGboost model demonstrated excellent discrimination and calibration, with AUCs of 0.823, 0.810, and 0.773 in training, internal, and external validation cohorts.
• The model showed stability and clinical utility in predicting tigecycline-associated hypofibrinogenemia.

Conclusions:

• Machine-learning models show significant promise for predicting tigecycline-associated hypofibrinogenemia.
• SHAP-based interpretation allows visualization of individualized risk for clinicians.
• This approach can facilitate tailored prophylactic strategies, enhancing patient safety and treatment outcomes.