Ying Zeng,1,* Hong Lu,1,* Sen Li,2,* Qun-Zhi Shi,1 Lin Liu,1 Yong-Qing Gong,1 Pan Yan1 

1Department of Pharmacy, The Affiliated Changsha Central Hospital, Hengyang Medical School, University of South China, Changsha, 410004, People’s Republic of China; 2Department of Pharmacy, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, 430030, People’s Republic of China

*These authors contributed equally to this work

Correspondence: Pan Yan, Department of Pharmacy, The Affiliated Changsha Central Hospital, Hengyang Medical School, University of South China, Changsha, People’s Republic of China, Email 2022050025@usc.edu.cn

Purpose: Drug-induced liver injury (DILI) is one of the most common and serious adverse drug reactions related to first-line anti-tuberculosis drugs in pediatric tuberculosis patients. This study aims to develop an automatic machine learning (AutoML) model for predicting the risk of anti-tuberculosis drug-induced liver injury (ATB-DILI) in children.
Methods: A retrospective study was performed on the clinical data and therapeutic drug monitoring (TDM) results of children initially treated for tuberculosis at the affiliated Changsha Central Hospital of University of South China. After the features were screened by univariate risk factor analysis, AutoML technology was used to establish predictive models. The area under the receiver operating characteristic curve (AUC) was used to evaluate model’s performance, and then the TreeShap algorithm was employed to interpret the variable contributions.
Results: A total of 184 children were enrolled in this study, of whom 19 (10.33%) developed ATB-DILI. Univariate analysis showed that seven variables were risk factors for ATB-DILI, including the plasma peak concentration (Cmax) of rifampicin, body mass index (BMI), alanine aminotransferase, total bilirubin, total bile acids, aspartate aminotransferase and creatinine. Among the numerous predictive models constructed by the “H2O” AutoML platform, the gradient boost machine (GBM) model exhibited the superior performance with AUCs of 0.838 and 0.784 on the training and testing sets, respectively. The TreeShap algorithm showed that Cmax of rifampicin and BMI were important features that affect the AutoML model’s performance.
Conclusion: The GBM model established by AutoML technology shows high predictive accuracy and interpretability for ATB-DILI in children. The prediction model can assist clinicians to implement timely interventions and mitigation strategies, and formulate personalized medication regimens, thereby minimizing potential harm to high-risk children of ATB-DILI.

Keywords: anti-tuberculosis drug-induced liver injury, children, retrospective study, automatic machine learning, gradient boost machine