To our knowledge, this study represents the first attempt to use local electronic health records to evaluate the prediction of DILI in an Asian population with tuberculosis, who are primarily Han Chinese. We observed a slight improvement in discriminatory ability for the ML model compared to the logistic model. Logistic regression offers better clinical generalizability but struggles to handle overfitting and missing variables, resulting in lower overall performance than expected. In contrast, both XGBoost and RF employ more advanced technology. XGBoost leverages gradient boosting to incrementally build weak learners and effectively capture nonlinear relationships with built-in regularization. On the other hand, RF, a bagging ensemble method, builds independent decision trees on random subsets of data, resulting in robust averaging but less explicit regularization. XGBoost excels at capturing complex nonlinear patterns and is well-suited for tasks with complex and dynamic interactions, such as predicting her DILI during tuberculosis treatment. Its training efficiency is also evident when processing large datasets. Although RF with robust averaging is suitable for further applications in diverse datasets, it can pose challenges in effectively capturing subtle nonlinear patterns among multiple explanatory variables.
Several previous studies have identified risk factors associated with DILI during tuberculosis treatment, including chronic liver disease, certain drug combinations, age, and various demographic characteristics. [25,26,27]. Lammert et al. [28] Our results suggested an increased risk of DILI in chronic liver disease patients presenting with NAFLD. Chan et al. [29] The addition of PZA to INH and RIF was shown to significantly increase the risk of hepatotoxicity. Hosford et al. [30] Through a systematic literature review, we established a significantly increased risk of hepatotoxicity in individuals aged 60 years and older. Abara et al. [2] They found that patients' lower body weight, HIV-1 co-infection, higher baseline ALP levels, and alcohol intake were risk factors. Therefore, in our model, we included enzyme levels, utilization of antituberculous drugs such as PZA, INH, and RIF, hepatoprotective agents such as silymarin and glycyrrhetinic acid, and demographic variables such as alcohol intake, age, gender, and education level. defined in advance. Ethnicity, occupation as a predictor. In the ultimate XGBoost model, the contribution weights of chronic liver disease, ALT, ALP, Tbil's ULN, and age were above 0.01, which is consistent with the findings of previous studies.
Currently, various predictive models for DILI are mainly working at the molecular level in the preclinical setting. [31]utilizes various artificial intelligence-assisted algorithms [32]. Minerali et al. [33] employed a Bayesian ML method, resulting in an AUROC of 0.81, sensitivity of 74%, specificity of 76%, and accuracy of 75%. Xu et al. [34] proposed a deep learning model that achieved 87% accuracy, 83% sensitivity, 93% specificity, and AUROC 0.96. Dominic et al.'s Bayesian prediction model [35] demonstrated balanced performance with accuracy of 86%, sensitivity of 87%, specificity of 85%, positive predictive value of 92%, and negative predictive value of 78%. At the clinical stage, only Zhong et al. Using a clinical sample of 743 TB cases, we introduced a single-tree XGBoost model with 90% precision, 74% recall, and 76% classification accuracy for DILI prediction. [36]. In our study, we leveraged local medical data and employed the XGBoost algorithm. The model showed 76% recall, 82% specificity, and 81% accuracy in predicting DILI status. Our approach was proven to be robust, with an average AUROC of 0.89 and AUPR of 0.75 at 10-fold cross-validation. During the clinical treatment phase, our model showed a high level of accuracy and interpretability.
The choice of cutoff in a DILI prediction model is very important and depends on the specific research goals and requirements. To improve understanding and prediction accuracy, various studies have investigated the optimal cutoff values for DILI prediction models. For example, a study focused on drug-induced liver tumors utilized the maximum Youden index to determine the ideal cutoff point. [37]. Another study aimed at predicting DILI and cardiotoxicity determined an optimal cutoff value of 0.4 using chemical structure and in vitro assay data. [38]. Similarly, a system called DILIps designed to predict his DILI in drug safety utilized ROC curves to select optimal cutoff values. [39]. Considering the unbalanced dataset of our study, we found that the precision-recall curve method seems to be more appropriate. Furthermore, considering the serious consequences of DILI, we recommend choosing a lower cutoff to maximize sensitivity, prioritizing the detection of DILI. Therefore, in our study, we selected the largest Youden index as the best cutoff.
However, the acceptance of ML in the medical community faces significant hurdles regarding interpretability, especially in settings where clinical decisions are paramount. In our study, we adopted the SHAP strategy to reveal the complex mechanism of the XGBoost model.
