Generative AI in Cancer Care: Mapping Review Finds Rapid Growth, Weak Evidence

A sweeping mapping review of 218 empirical studies has found that large language models and other generative AI tools are spreading rapidly through cancer care research, yet almost all of the evidence behind them measures correctness rather than whether they actually improve clinical outcomes or patient safety. The review, published in BMC Artificial Intelligence, followed PRISMA reporting guidelines and searched five databases through June 2025 to characterize how these tools are developed, evaluated, and implemented in oncology. Its authors conclude that the field has entered a rapid prototyping era anchored to general-purpose, closed models, with validation focused overwhelmingly on accuracy and content quality rather than clinical utility.

The growth curve is striking. Only 24 eligible articles were published in 2023, but that figure jumped to 125 in 2024, a 5.2-fold increase, and 69 more eligible articles appeared between January and June 2025 alone. The activity is heavily concentrated in wealthy countries: 74 studies (33.9 percent) came from the United States, 39 (17.9 percent) from China, and 24 (11 percent) from Germany, reflecting strong engagement across North America, Europe, and East Asia while leaving much of the world’s health systems and languages untested.

Functionally, the tools cluster into a few categories. Clinical decision support systems dominated, accounting for 61.5 percent of studies (n=134), followed by patient-facing chatbots and conversational agents at 17.4 percent (n=38) and patient education tools at 11.9 percent (n=26). Information extraction and summarization tools made up 5.5 percent, and documentation, screening, and communication aids just 3.7 percent. In terms of the decisions supported, diagnosis and screening tasks led with 118 studies, roughly 54 percent of all tools, while 39 percent (n=85) supported treatment selection or planning. Follow-up decision aids (12 percent, n=26) and prognostic tools (9 percent, n=19) were far less common, and survivorship support was rarely addressed.

The intended users were mostly clinicians. Healthcare professionals were the target group in about 73 percent of studies (n=159), with oncologists the primary users in 39 percent and general clinicians in 22.5 percent. Still, patient-facing applications were prevalent overall, making up 55 percent of studies (n=120), including tools for patients (50.5 percent) and caregivers (3.7 percent). Breast cancer was the most studied cancer type (17 percent), followed by prostate (13 percent), head-and-neck (11 percent), and lung cancers (9 percent), with 14 percent of studies covering multiple or general cancer types. Most studies did not specify cancer stage.

The underlying technology reveals a strong reliance on convenience. ChatGPT-family models, including GPT-3.5, GPT-4, and GPT-4o, were used in 78 percent of studies (n=170), with Google Bard/Gemini appearing in about 20 percent, Claude in roughly 10 percent, and Microsoft Copilot in 8 percent. Only 10 percent of articles explicitly described adapting the models, through techniques such as retrieval-augmented generation, chain-of-thought prompting, or custom GPTs, and a mere 3 percent (n=6) fine-tuned models for the oncology domain. Nearly all articles relied on natural-language input, and 93 percent used the public web or chatbot version of the tools; only 7 percent (n=15) were embedded in electronic health records.

Evaluation methods were similarly narrow. The most common designs were expert or Likert grading of answers, guideline concordance tests, and benchmark question-and-answer experiments using fixed question banks, along with retrospective dataset analyses and occasional clinician user studies. Prospective and real-world evaluations remained uncommon. Across the included studies, measured accuracy ranged from 55 to 97 percent, improving when clinical context or multimodal input was introduced, with image-plus-history cases reaching 73 to 93 percent. ChatGPT-4 consistently outperformed other models, achieving 75 to 97 percent correctness in most benchmark tasks compared with 60 to 70 percent for GPT-3.5 and 40 to 60 percent for Bard, Bing, or Gemini. Notably, some domain-specific fine-tuned models, such as RadOnc-GPT and LilyBot, sometimes exceeded the GPT baselines in correctness or time efficiency.

Concordance with clinical guidelines ranged from 33 to 87 percent, again with ChatGPT-4 leading, and GPT-4 achieved substantial alignment with tumor board and reference decisions (kappa of 0.72 to 0.88). Most studies reported high reproducibility above 90 percent. Readability, however, was a persistent weakness: outputs generally corresponded to a college reading level (FKGL around 11 to 13), although prompting to a sixth-grade level improved comprehension scores. Quality instruments such as DISCERN and PEMAT indicated moderate to good quality, with understandability ranging from about 65 to 85 percent but actionability between only 40 and 60 percent. Clinicians rated GPT-4 outputs around 4 to 4.5 out of 5 for accuracy, clarity, and comprehensiveness, while patients highlighted ease of understanding and empathy.

The reported challenges form a sobering counterweight to the enthusiasm. Common issues included hallucinations, false citations, and outdated information stemming from static training data and limited access to proprietary oncology databases. Many tools lacked explainability, offering little insight into their data sources or reasoning, which undermined verification and clinician confidence. Patient-facing outputs often exceeded recommended readability levels, and ethical concerns arose around privacy, liability, and the absence of quality controls in web-based tools. Implementation obstacles included poor EHR integration, dependence on prompt quality, inconsistent reproducibility, and the lack of expert oversight. Oncology adds particular stakes, the authors note, including nuanced staging schemas, fast-moving guidelines, and high-risk therapy decisions.

Perhaps the most consequential gap is what was not measured. Few studies examined clinical, behavioral, or implementation outcomes such as decisional conflict, adherence, knowledge gain, clinician cognitive load, decision time, cost, or equitable access. The review’s overall stance analysis found that 51.4 percent of articles expressed balanced, cautious views of these tools, 39.6 percent were supportive, and 8.7 percent were critical. The authors caution that their findings may be overselling the potential of LLMs and generative AI, and that the evolving nature of the models makes outputs difficult to control and results hard to reproduce. Publication bias may also inflate perceived progress, since favorable or novel findings are more likely to be published than negative ones.

The review has its own limitations, which the authors spell out. By design, it excluded standalone predictive, classification, or detection models, so observations about model adaptation apply specifically to user-facing generative decision-support systems rather than oncology AI broadly. The exclusion of non-English literature may have omitted valuable studies, and the heterogeneity of evaluation metrics limited cross-study comparison. Because a mapping review characterizes a field rather than estimating pooled effects, no conclusions about intervention effectiveness can be drawn. The authors call for future research prioritizing context-specific model adaptation, workflow integration, equitable language accessibility, and rigorous mixed-method evaluations, including pragmatic studies embedded in actual clinical workflows, to determine whether these systems can mature from promising demonstrations into trustworthy, clinically meaningful tools for cancer care.

Subject of Research: A mapping review of empirically tested large language model and generative AI tools supporting decision-making in cancer care

Article Title: The use of Large Language Models and other Generative AI tools to support decision-making in cancer care: a mapping review

Article References: Elkefi, S., Scheinmann, R., Hao, D., Bhar, S., Tounsi, A., & Feiner, S. (2026). The use of Large Language Models and other Generative AI tools to support decision-making in cancer care: a mapping review. BMC Artificial Intelligence, 2(1), Article 7. https://doi.org/10.1186/s44398-026-00024-x

Image Credits: AI Generated

DOI: 10.1186/s44398-026-00024-x

Keywords: large language models, generative AI, cancer care, clinical decision support, oncology, mapping review, ChatGPT, GPT-4, patient-facing chatbots, hallucination, EHR integration, guideline concordance