As cancer medicine becomes increasingly data-intensive, researchers explore how generative AI could help oncologists interpret genomic mutations, identify clinical trials, and synthesize complex patient data while maintaining strict human oversight.

In a recent narrative review published in the Journal of Hematology & Oncology, researchers synthesized recent literature investigating the integration of generative artificial intelligence (AI) tools into clinical precision oncology practice. The review focused on how these tools can support oncologists in interpreting genetic mutations, including variants of uncertain significance and other complex genomic alterations that require extensive literature cross-referencing.

Furthermore, they evaluate the abilities of large language models (LLMs) to rapidly screen and match patients to clinical trial eligibility and to aid oncologists in drafting detailed reports of imaging findings, pathology results, and integrated tumor phenotypes that combine genomic, imaging, and clinical data. Finally, to ensure that these computational tools are leveraged in safe and responsible clinical environments, the authors propose novel AI operational frameworks that prioritize continuous human oversight and enable the grounding of model outputs in current medical knowledge through retrieval-based systems.

Precision oncology and the growing complexity of cancer data

Precision oncology is a relatively novel approach to cancer care that seeks to revolutionize oncology through personalized, targeted therapies tailored to a patient’s tumor-specific molecular profile.

Since their conceptualization and introduction in the early 2000s, these integrative strategies have produced major advances in certain cancers through biomarker-guided therapies, improving patient outcomes in selected clinical contexts compared with traditional “one-size-fits-all” chemotherapeutic interventions.

A growing body of literature, however, emphasizes that the volume and complexity of data generated by modern next-generation sequencing (NGS) technologies, combined with the need for clinicians to cross-reference these datasets against extensive patient electronic health records (EHRs) and rapidly expanding biomedical literature, may overwhelm human oncologists and degrade patient care.

Artificial intelligence emerges as a tool to help manage clinical data overload

The advent of the “AI age” is suggested to partially address these human limitations. Artificial intelligence (AI) models are capable of ingesting and processing extensive, multimodal datasets, helping clinicians synthesize information from genomic data, clinical records, imaging, and scientific literature. Unfortunately, previous research cautions against persistent limitations of incorporating generative AI in high-stakes healthcare environments.

The most frequently cited of these risks is “AI hallucination,” in which the model confidently invents incorrect information to meet user requirements. Such errors may include fabricated references, misinterpreted genomic variants, or incomplete summaries of clinical evidence, highlighting the importance of careful clinical validation and supervision of AI systems in precision cancer care.

Review evaluates emerging AI tools across oncology workflows

The present review aimed to address this urgent and critical oncological need by synthesizing recent literature to map the current medical AI landscape and advise oncologists on recent technological advancements in the field, whilst simultaneously underscoring AI’s pitfalls and the need for constant human supervision.

The review surveyed peer-reviewed publications across several online databases, including PubMed, Web of Science, and Scopus, and cross-referenced these publications with official regulatory documents and medical device clearances issued by the U.S. Food and Drug Administration (FDA).

Review analyses evaluated the real-world performance of deep generative foundation models, specifically LLMs and vision-language models (VLMs), focusing on AI’s ability to interpret molecular and biomarker-derived data and translate their findings into clinically relevant summaries and decision-support information.

Furthermore, the review investigated AI’s efficacy in screening patients for eligibility in clinical trials and its capacity to leverage multimodal datasets to draft comprehensive reports that integrate imaging, pathology, genomic, and clinical information into structured tumor phenotype descriptions.

AI models demonstrate strong performance in clinical trial matching and imaging analysis

The review presents several scientifically validated examples that underscore the clinical benefits of using AI tools in precision oncology. For example, the review of TrialGPT found that the model, designed to assess patient suitability to ongoing clinical trials, achieved strong agreement with expert assessments in the evaluated study (87.3% accuracy) while reducing the processing time by an average of 42.6%.

Similarly, Flamingo-CXR (a novel VLM) was found to match or exceed the performance of board-certified radiologists in 94% of chest X-ray cases with no clinically relevant findings, and to produce diagnostic reports equal to or superior to those of human experts in 77.7% of evaluated inpatient and outpatient cases.

Human oversight remains essential due to risks of AI errors

However, the review also highlighted critical vulnerabilities in the unsupervised use of AI in healthcare settings. In radiology imaging, for example, clinically significant interpretation errors were reported in both AI- and human-generated reports in 24.8% of evaluated cases, with additional errors occurring exclusively in AI- or human-generated reports, underscoring the importance of maintaining human oversight when deploying these systems in clinical environments.

To mitigate these inherent risks, reviewers advocate for “Human-in-the-Loop” (HITL) workflows, which require human experts to review AI outputs before they are clinically implemented.

Additionally, they emphasize the use of Retrieval-Augmented Generation (RAG), a technique that forces the AI to frequently retrieve information directly from current guidelines, biomedical literature, and curated medical databases rather than relying on its base memory.

AI should assist oncologists rather than replace clinical decision-making

The present review highlights AI’s benefits as an oncologist’s assistant, demonstrating how these models can rapidly synthesize complex data and draft detailed diagnostic reports, thereby helping human experts make faster, more informed decisions.

However, it simultaneously cautions that AI’s widespread clinical adoption in personalized cancer care should not be rushed, emphasizing the need for established medical data privacy standards, addressing persistent demographic biases in training datasets, and the critical requirement for continuous human oversight.

In conclusion, the review posits that current AI implementations cannot be given autonomous decision-making responsibilities but should be viewed as tools that aid human oncologists in treating cancer.