Risk prediction models for anti-tuberculosis drug-induced liver injury show pooled AUC of 0.81Risk Prediction Models Help Identify Liver Damage from Tuberculosis Drugs

ObjectiveThis study systematically evaluates the construction methods, predictive factors, and performance of risk prediction models for anti-tuberculosis drug-induced liver injury (ATB-DILI). It aims to furnish evidence-based evidence for early detection of East Asian individuals at high risk of ATB-DILI and the development of personalized clinical treatment plans.MethodsA comprehensive literature search was conducted to identify studies related to risk prediction models for ATB-DILI. Key data, including study population, sample size, model construction algorithms, predictive factors, and model discrimination, were extracted. Qualitative synthesis and systematic analysis were then performed.ResultsA total of 26 studies comprising 31 risk prediction models (cumulative sample size approximately 41,734 cases) were included. Twenty studies reported internal validation and six reported external validation. Model discriminative ability varied from 0.624 to 1.0, indicating at least modest-to-excellent discriminatory ability. Prediction model Risk Of Bias ASsessment Tool (PROBAST) assessment identified only one study at low risk of bias, with all remaining studies rated high risk of bias. Meta-analysis yielded a pooled AUC of 0.81 for ATB-DILI prediction models. History of liver disease, extrapulmonary tuberculosis, age ≥60 years, alcohol consumption, concomitant medication use, retreatment, elevated AST, diabetes, and smoking were associated with higher ATB-DILI risk (all P < 0.05), whereas uric acid showed an inverse association (P < 0.05). Importantly, the observed association for non-use of hepatoprotective agents cannot confirm a preventive effect; these factors should inform risk stratification and intensified monitoring rather than guide prophylactic hepatoprotective therapy.ConclusionPublished risk prediction models for ATB-DILI among East Asian populations were systematically appraised. These models yielded moderate-to-good discriminative performance yet showed substantial heterogeneity and high risk of bias, mainly due to defective missing-data processing, univariate-based predictor selection, insufficient outcome events, and absent external validation. Such methodological shortcomings may overestimate AUC values, especially for retrospective non-validated models. Existing ATB-DILI prediction models are thus of uncertain reliability and generalizability. Future large-sample multicenter prospective studies with standardized outcomes and external validation are required to develop robust clinically applicable models.Systematic review registrationhttps://www.crd.york.ac.uk/PROSPERO/view/CRD420261468400, identifier CRD420261468400.