Interpretable machine learning in predicting drug-induced liver injury among tuberculosis patients: model development...

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Pharmacovigilance

Background:

• Drug-induced liver injury (DILI) is a significant concern during tuberculosis (TB) treatment.
• Predictive models are needed to identify patients at high risk of DILI.

Purpose of the Study:

• To develop and validate an interpretable prediction model for DILI in TB patients.
• To compare the performance of XGBoost, Random Forest, and LASSO logistic models.

Main Methods:

• Utilized a dataset of 7,071 TB patients from Ningbo City.
• Developed and compared XGBoost, RF, and LASSO logistic models.
• Employed SHAP for model interpretability and identified key predictive variables.

Main Results:

• The XGBoost model achieved superior performance with an AUROC of 0.89 and AUPR of 0.75.
• Key predictors for DILI included pre-existing liver conditions (DIH, FLD), age, ALT, and Tbil.
• SHAP analysis elucidated the contribution of these variables to DILI prediction.

Conclusions:

• XGBoost offers improved predictive accuracy for DILI in TB patients compared to RF and LASSO.
• SHAP enhances the clinical interpretability and applicability of the DILI prediction model.
• Further research should focus on external validation and feature integration.