Artificial Intelligence Approaches for Prediction and Detection of Immune-Related Adverse Events with Immune Checkpoint Inhibitor Cancer Therapy: A Narrative Review

Immune checkpoint inhibitors (ICIs) have translated the scope of cancer therapy, but their immune-restorative mechanism can lead to immune-related adverse events (irAEs), which can affect several organs with varying severity. Early identification of patients at risk of irAEs is prudent to inform clinical decisions in precision oncology and remains a challenge as current clinical and biomarker methods still lack predictive accuracy. Artificial Intelligence (AI) presents a promising strategy for improving early detection, risk stratification as well as monitoring of irAEs. This narrative review summarizes the current AI modalities for detecting and predicting irAEs risk occurring with ICI therapy, including clinical Machine Learning models, radiomics-based approaches, natural language processing (NLP) systems and the integration of modalities with multimodal AI frameworks. Clinical Machine Learning models demonstrate moderate predictive performance whereas radiomics-derived modeling appears promising for pneumonitis. NLP and language models have achieved higher accuracy for retrospective irAEs detection. Multimodal AI applications offer theoretical potential through diverse data integration that captures the complex biology of irAEs; however, current evidence is limited. All these modalities face limitations of inadequate sample sizes, retrospective design, heterogeneous outcome definitions, class imbalance, and insufficient external validation. AI-based models have significant potential for personalized immunotherapy monitoring but require prospective multicenter validation, standardized datasets and clinically interpretable frameworks prior to implementation. Future advances in multimodal modeling can also enable precise prediction and early detection of irAEs, ultimately improving the safety and effectiveness of cancer immunotherapy.