Explainable Artificial Intelligence for Predicting Cancer Treatment Toxicity in Real-World Clinical Practice

Abstract

Cancer treatment–related toxicities remain a major challenge in oncology, frequently compromising patient quality of life, treatment adherence, therapeutic efficacy, and overall survival. The increasing availability of real-world clinical data (RWD), combined with advances in artificial intelligence (AI) and machine learning (ML), has created new opportunities for the early prediction of treatment-related adverse events and personalized toxicity risk assessment. This review provides a comprehensive overview of AI-driven approaches for predicting cancer treatment toxicity using real-world clinical data. It examines the application of conventional machine learning, deep learning, and emerging foundation models across diverse data sources, including electronic health records, genomic and molecular datasets, medical imaging, laboratory parameters, wearable device data, and patient-reported outcomes. Machine learning models have demonstrated encouraging predictive performance for identifying treatment-related adverse events. Among the commonly used algorithms, random forest has been the most frequently applied, followed by support vector machines, XGBoost, decision trees, and LightGBM. A recent meta-analysis reported a pooled sensitivity of 0.65, specificity of 0.89, and an area under the receiver operating characteristic curve (AUC) of 0.8069, highlighting the potential of AI-based models for clinical toxicity prediction. Recent advances in transformer-based architectures, multimodal learning, and large language models have further improved predictive accuracy and model generalizability across multiple cancer types, treatment modalities, and healthcare settings. Foundation models trained on large-scale multimodal real-world datasets are emerging as promising clinical decision-support tools capable of integrating heterogeneous patient information for individualized toxicity prediction. AI-driven toxicity prediction has the potential to transform precision oncology by enabling proactive patient monitoring, individualized treatment planning, dose optimization, early intervention, reduced treatment-related hospitalizations, and improved clinical outcomes. Despite these advances, challenges related to data quality, model interpretability, external validation, fairness, and seamless clinical implementation remain significant barriers to widespread adoption. This review summarizes current evidence, discusses emerging technologies, identifies existing limitations, and highlights future directions for integrating AI-driven toxicity prediction into routine oncology practice.