Glucocorticoids are widely used in managing severe drug-induced liver injury (DILI) to enhance clinical improvement and reduce the length of hospital stay, yet they may elevate susceptibility to infections. This work aimed to build a model capable of forecasting infection following glucocorticoid administration in individuals with DILI. A retrospective review was carried out on patients diagnosed with severe DILI who received glucocorticoid treatment at the Fifth Medical Center of the Chinese People’s Liberation Army from 2017 to 2024. Eight machine-learning approaches were developed and assessed: random forest, support vector machine, generalized linear model, gradient boosting machine, least absolute shrinkage and selection operator, XGBoost, K-nearest neighbor classification, and an artificial neural network. The most effective model was examined using decision curve analysis, calibration plots, ROC curves, and Shapley Additive Explanations.
Of the eight algorithms, the gradient boosting machine yielded the strongest performance, achieving an area under the ROC curve of 0.981 for the validation cohort and 0.928 for the test cohort, along with the lowest residuals. Its clinical utility was further supported by decision curve analysis and calibration plots. Among the predictive features, globulin (GLO) stood out, showing markedly lower concentrations in infected individuals than in non-infected patients (p < 0.001). Those whose GLO values before treatment were under 20 g/L demonstrated an infection rate of 41.1%, while individuals with post-treatment GLO below 21.5 g/L had an even higher infection rate of 82.3%. The early-warning model presented here offers practical value for guiding glucocorticoid therapy in severe DILI. Tracking GLO level fluctuations may represent a straightforward and efficient method for identifying patients at increased risk of infection.
