Development of a Predictive Model for Classifying Immune Checkpoint Inhibitor‐Induced Liver Injury Types

Abstract

ABSTRACT Aims Immune checkpoint inhibitors (ICIs) have transformed cancer therapy; however, they are associated with ICI‐induced liver injury (ICI‐LI), which manifests as hepatocellular, mixed, or cholestatic patterns with variable treatment responses. This study aimed to develop and validate a predictive model to identify ICI‐LI type using clinical data available at ICI initiation. Methods A retrospective analysis of 297 patients with ICI‐LI was conducted. Baseline clinical data were analyzed using univariate and multivariate logistic regression to predict ICI‐LI types in the training and validation cohorts. A predictive model was developed and validated using receiver operating characteristic (ROC) curve analysis. Results Multivariate analysis in the training cohort identified male sex (odds ratio [OR]: 3.33, 95% confidence interval [CI]: 1.57–7.06, p = 0.002), serum albumin levels (OR: 0.42, 95% CI: 0.19–0.91, p = 0.027), and serum alanine aminotransferase (ALT) levels (OR: 0.97, 95% CI: 0.94–0.99, p = 0.015) as significant predictors, along with ICI regimen types selected using the Akaike information criterion. The logistic regression model, expressed as p = 1/{1 + (−(5.02 + 1.20 × (sex [F:0, M:1])) − 0.87 × albumin [g/dL] − 0.03 × ALT [U/L] − 0.9 × (drug [non‐anti‐cytotoxic T lymphocyte antigen 4 (CTLA‐4) related regimen:0, anti‐CTLA‐4 related regimen:1]))}, achieved an area under the ROC (AUROC) of 0.73 (95% CI: 0.63–0.82) in the training cohort. At a cut‐off of 0.86, the sensitivity was 60.3%, specificity 74.4%, positive predictive value 92.3%, and negative predictive value 26.9%. In the validation cohort, the AUROC was 0.752 (95% CI: 0.476–1.00). Conclusion This predictive model demonstrates its utility in classifying ICI‐LI types.

Keywords

• immune checkpoint inhibitor‐induced liver injury
• immune checkpoint inhibitors
• immune‐related adverse events
• predictive model
• treatment management