Analysis of risk factors for immune checkpoint inhibitor-associated liver injury: a retrospective analysis based on...

Immune-mediated hepatotoxicity (IMH) induced by immune checkpoint inhibitors (ICIs) can lead to fatal outcomes. Exploring the risk factors associated with IMH is crucial for the early identification and management of immune-related adverse events (irAEs).

Screening IMH-influencing factors by applying meta-analysis to clinical research data. Utilizing FAERS data, ICIs-related IMH prediction models were developed using two types of variables (full variables and optimal variables screened by univariate logistic regression) and nine machine learning algorithms (logistic regression, decision tree, random forest, gradient boosting decision tree, extreme gradient boosting, K-Nearest Neighbor, bootstrap aggregation, adaptive boosting, and extremely randomized trees). Comparing the nine machine learning algorithms and screening the optimal model while using SHAP (SHapley Additive exPlanations) analysis to interpret the results of the optimal machine learning model.

A total of 17 studies (10,135 patients) were included. The results showed that ICIs combination therapy (OR = 5.10, 95% CI: 1.68-15.48) and history of ICIs treatment (OR = 3.58, 95% CI: 2.08-6.14) were significantly associated with the risk of all-grade IMH. Patients aged 56-63 years (MD = - 5.09, 95% CI: - 9.52 to - 0.67) were significantly associated with the risk of ≥ grade 3 IMH. The liver adverse reaction prediction model included a total of 51,555 patients from the FAERS database, of which 4607 cases were liver adverse reactions. Univariate logistic regression analysis ultimately screened eight optimal variables, with females, report areas, cancer type, ICIs drug type, concomitant autoimmune disease, the concomitant use of anti-hypertension drugs, and the concomitant use of CTLA-4 inhibitors or targeted therapy drugs being significant influencing factors. The performance of the model after the variables were screened by univariate logistic regression was slightly worse than that of the model with full variables. Among the best-performing liver adverse reaction prediction models was GBDT (training set AUC = 0.82, test set AUC = 0.79). The top 3 key predictors in the GBDT model were report areas, disease type, and ICIs drug type.

In clinical studies, we found that age between 56 and 63 years, ICIs combination therapy, and history of ICIs treatment were significantly associated with an increased risk of IMH. In the FAERS database, we observed that females, report areas, cancer type, ICIs drug type, concomitant autoimmune disease, the concomitant use of anti-hypertension drugs, and the concomitant use of CTLA-4 inhibitors or targeted therapy drugs may be potential risk factors for ICIs-related hepatic irAEs. The predictive model for ICIs-related liver adverse reactions established in this study has good performance and potential clinical applications.