TY  - JOUR
T1  - Early Prediction of Non-recovery in Drug-induced Liver Injury by Integrating Genetic Variants and Clinical Variables: An Interpretable Machine Learning Approach 
AU  - Jiang, Jingjing
AU  - Lou, Weiwei
AU  - Li, Qing
AU  - Li, Ziqiang
AU  - Lou, Weiqian
AU  - Zhu, Xichen
AU  - Xie, Qing
AU  - Lai, Rongtao
JF  - Journal of Clinical and Translational Hepatology
VL  - 
IS  - 000
SN  - 2310-8819
SP  - 
EP  - 
Y1  - 2026-08-05
DO  - 10.14218/JCTH.2026.00447
UR  - https://www.xiahepublishing.com/2310-8819/JCTH-2026-00447
AB  - Background and Aims 
Early predictors of 6-month non-recovery in drug-induced liver injury (DILI) remain limited. Genetic variants are stable host characteristics that may complement baseline clinical variables. We aimed to develop and validate an interpretable, clinical-genetic machine learning model for predicting 6-month non-recovery in patients with DILI.
Methods
This retrospective, single-center study included 338 patients with DILI, who were classified as recovered (n = 171) or non-recovered (n = 167) at 6 months. Candidate single-nucleotide polymorphisms and baseline clinical variables were collected during initial hospitalization. Features were selected using complementary screening approaches. Multiple machine learning models were developed and compared. Model discrimination, calibration, clinical utility, the incremental value of genetic predictors, and interpretability using SHapley Additive exPlanations (SHAP) were assessed.
Results
Five predictors were consistently retained for model development: rs72631567, rs28521457, alanine aminotransferase, monocyte percentage, and low-density lipoprotein. Among the candidate algorithms, the light gradient boosting machine model showed the best performance, with area under the receiver operating characteristic curve (AUC) values of 0.92 (95% confidence interval [CI] 0.89–0.95) in the training set and 0.81 (95% CI 0.70–0.91) in the validation set. The model showed acceptable calibration and favorable decision-curve performance. In ablation analysis, the clinical-only model showed limited discrimination (AUC 0.57, 95% CI 0.43–0.71). SHAP analysis identified rs72631567 as the most influential predictor.
Conclusions
An interpretable model that integrates host genetic variants with baseline clinical variables demonstrated good internal performance for early prediction of 6-month non-recovery in DILI. These findings support external validation of genotype-informed risk stratification in patients with DILI.