Machine learning (ML) and deep learning (DL) models applied to electronic health records (EHRs) have substantial potential to improve oncology care across diagnosis, prognosis, treatment selection, and trial recruitment. However, opacity of many high-performing models limits clinician trust, regulatory acceptance, and safe deployment. Explainable artificial intelligence (XAI) methods aim to make model behavior understandable and actionable in clinical contexts. The present perspective summarizes current XAI approaches applied to EHR-based oncology tasks, identifies key challenges in evaluation, reproducibility, clinical utility, and equity, and proposes pragmatic recommendations and research directions to accelerate safe adoption in oncology. Common XAI categories used with EHR data include feature importance/interaction methods, intrinsically interpretable models, attention mechanisms, dimensionality reduction, and knowledge distillation or rule extraction. Tree-based models with SHapley Additive exPlanations (SHAP) explanations dominate recent EHR studies. Other interpretable strategies, such as generalized additive models and rule sets, appear in settings where transparency is prioritized. Gaps include inconsistent reporting, scarce formal evaluation of explanations for clinical utility, limited reproducibility for data and code availability, inadequate external validation, and insufficient consideration of fairness and equity that these issues are particularly important in oncology, where heterogeneity and stakes are high. Overall, integrating XAI with EHR-driven oncology models is promising but underdeveloped, which requires further progress by multi-stakeholder evaluation frameworks, reproducible pipelines, prospective and multicenter validations, and equity-aware design. The field should prioritize clinically meaningful explanations beyond ranking features and study how explanations affect clinician decision-making and patient outcomes.
Yang et al. (2026) studied this question.