OBJECTIVE A machine learning algorithm was used to construct a risk prediction model for childhood antituberculosis drug-induced liver injury(ATB-DILI), providing a new method for accurate prediction of ATB-DILI in clinical pediatric tuberculosis patients.

METHODS Pediatric patients diagnosed with tuberculosis in Children's Hospital of Chongqing Medical University from January 2013 to December 2022 were selected for the study. Univariate and LASSO regression were used to screen the characteristic variables. They were randomly divided into training set(1 957 cases) and test set (839 cases) in the ratio of 7∶3. The training set was used for risk prediction model construction and parameter adjustment, and the test set was used to validate the model performance. Extreme gradient boosting(XGBoost), adaptive boosting, light gradient boosting machine, and random forest machine learning algorithms to construct the prediction model. The model performance was evaluated by the area under the curve(AUC), accuracy, precision, recall and F1 score of the subjects' work characteristics, and the Shapley additive explanation(SHAP) algorithm was used to perform interpretive analysis of the optimal model to quantify and visualise the presentation of the complex relationship between the risk factors and the prediction results, and to increase the interpretability of the model. Nomogram make the prediction results more intuitive.

RESULTS A total of 2 796 tuberculosis patients were included and the incidence of ATB-DILI was 5.47%. The XGBoost model had the best overall predictive performance with AUC(0.881), precision(0.951), accuracy(0.981), recall(0.956) and F1 score(0.968). The importance of clinical features in the XGBoost model was ranked based on the SHAP algorithm, in the order of treatment history, baseline value of alanine aminotransferase, baseline value of gamma-glutamyl transferase, number of days of hospitalisation, number of days of isoniazid treatment, total cumulative dose of isoniazid, baseline value of direct bilirubin, and number of days of ethambutol treatment. An in-depth analysis of how the characteristic variables in the XGBoost model affected the predicted outcomes revealed that tuberculosis retreatment, high baseline values of alanine aminotransferase, high baseline values of gamma-glutamyl transferase, high baseline values of direct bilirubin, prolonged hospitalisation days, long days of ethambutol treatment, and high cumulative total dose of isoniazid had a positive impact on the predicted outcomes, tending to favour the occurrence of ATB-DILI.

CONCLUSION The risk prediction model for ATB-DILI in children was constructed based on the XGBoost algorithm performed optimally. The SHAP algorithm provides an explicit interpretation of the model and clarifies that tuberculosis resumption, high baseline values of alanine aminotransferase, high baseline values of gamma-glutamyl transferase, high baseline values of direct bilirubin, prolonged days of hospitalisation, days ≤45 of isoniazid treatment, high cumulative total dose of isoniazid and long days of ethambutol treatment are the risk factors for children's risk factors for ATB-DILI in patients with tuberculosis. Nomogram make the predictive results easy to understand and apply, and helps in early clinical identification and prevention of ATB-DILI in children.