Clinical performance and readiness of artificial intelligence models for predicting immune-related adverse events in patients receiving immune checkpoint inhibitors: A systematic review.

e13683 Background: Immune checkpoint inhibitors (ICIs) have transformed cancer treatment but are frequently complicated by immune-related adverse events (irAEs), which may result in treatment interruption, hospitalization and mortality. Reliable pre-treatment tools to identify patients at higher risk for irAEs, particularly checkpoint inhibitor pneumonitis (CIP) remain limited. Artificial intelligence (AI)-based prediction models have emerged but their clinical performance and readiness for implementation are unclear. Methods: We systematically searched PubMed, Embase, Scopus and Web of Science (inception-January 2025) for studies developing AI-based models to predict irAEs or CIP in patients receiving ICIs. Eligible studies included radiomics, deep learning, natural language processing and clinical machine-learning approaches. Two reviewers independently extracted data on cancer type, toxicity endpoint, modality, algorithm, validation strategy and model performance. Study design and reporting characteristics were assessed. Due to substantial methodological heterogeneity, results were synthesized descriptively. Results: Fifteen studies met inclusion criteria, including mixed solid tumor cohorts and lung-predominant populations. Predictive targets included any-grade irAEs and organ-specific toxicities, most commonly pneumonitis. Model approaches ranged from clinical and electronic health record–based models to CT- and PET/CT-based radiomics and multimodal imaging-clinical frameworks. Reported discriminatory performance varied widely, with AUC values ranging from 0.64 to > 0.90 across model types. The most consistent performance was observed in deep learning and multimodal radiomics-clinical models for pneumonitis, with AUCs frequently between 0.75 and 0.92, whereas several clinical-only models demonstrated more modest performance. Most studies were retrospective and single-center, used heterogeneous irAE definitions and lacked standardized imaging protocols. External validation and calibration assessment were uncommon. Conclusions: AI-based models for predicting irAEs in ICI-treated patients demonstrate promising performance, particularly for pneumonitis using radiomics and multimodal approaches. However, current evidence is limited by heterogeneity, scarce external validation and uncertain generalizability. Prospective, multi-institutional validation and standardized outcome definitions are needed before AI-driven irAE risk prediction can be integrated into routine oncology practice to guide surveillance and patient counseling.