Large Language Models in Oncology: A Paradigm Shift from Discriminative Pattern Recognition to Generative Clinical Intelligence

1. Introduction

On November 30, 2022, OpenAI made ChatGPT available to the general public. Within five days, the platform had registered one million users; within two months, one hundred million [1]. No consumer technology in recorded history had achieved comparable adoption velocity, a phenomenon that prompted researchers to characterize the event as a tipping point for artificial intelligence in medicine [1]. For oncology, a specialty that had already demonstrated meaningful engagement with earlier generations of artificial intelligence, this moment signaled a categorical change in what machines were capable of doing with clinical information: a transition from narrow perceptual discrimination to broad generative reasoning.

The preceding decade had delivered genuine advances in discriminative artificial intelligence: systems trained to identify malignant lesions on mammography, grade tumor histopathology, or stratify molecular risk profiles from genomic data (2, 3). These systems were, by design, bounded. Each was optimized for a single perceptual task and could not extend beyond it. LLMs belong to an architecturally distinct paradigm. Built on the transformer, introduced in 2017 and described in the foundational work of Vaswani and colleagues, and trained on corpora of text at a scale measured in hundreds of billions of tokens, LLMs acquire emergent capacities that were not explicitly programmed, including multi-step reasoning, analogical inference, and instruction-following across heterogeneous domains (4, 5). The resulting systems can synthesize a complex clinical narrative, navigate a guideline document, translate a technical report into patient-accessible language, and engage in iterative dialogue, all within a single architectural framework (6, 7).

Successive model families, from GPT-4 through GPT-4o as described by OpenAI, Google's Gemini series as characterized by the Gemini Team, Anthropic's Claude series, and the open-source Llama lineage, have each produced measurable improvements in accuracy, context length, and domain-specific competence (8, 9). By early 2026, a living systematic review documented more than 4,600 peer-reviewed publications on LLMs in clinical medicine published between January 2022 and September 2025, representing approximately 3.2 papers per day [10]. Within oncology, systematic reviews, scoping reviews, and institutional guidance documents have proliferated at a pace that challenges synthesis (11, 12, 13). This review is organized around a three-tier framework of clinical readiness and a set of cross-cutting technical and ethical themes identified as insufficiently addressed in prior reviews.

Review methodology note. This is a narrative review. Literature was identified through systematic searches of PubMed, MEDLINE, Embase, and the Cochrane Library, supplemented by targeted searches of Google Scholar and preprint servers (arXiv, medRxiv), institutional repositories, and regulatory authority websites. Search terms included combinations of: “large language models,” “generative AI,” “oncology,” “cancer,” “retrieval-augmented generation,” “multimodal,” “clinical decision support,” “hallucination,” “patient education,” “health literacy,” “algorithmic bias,” and “software as a medical device.” The search period covered November 2022 through March 2026. Priority was given to peer-reviewed original studies, systematic reviews, and meta-analyses; where prospective peer-reviewed evidence was unavailable, high-impact technical preprints and institutional guidance documents were included for emerging domains. Preprint citations are identified by their server of origin (arXiv or medRxiv) and should be interpreted with appropriate caution pending formal peer review.

Table 1. Three-Tier Framework of LLM Clinical Readiness in Oncology

Tier	Clinical Domain	Key Evidence
Tier 1: Established Performance	Literature synthesis and evidence mining; clinical documentation; radiology and pathology report simplification; patient education and plain-language communication	Van Veen et al. [40]: 10-fold time reduction in summarization; Ferber et al. [32]: RAG-GPT-4 accuracy 57% to 84%; Kwok et al. [49]: reading level 9.6 to 14 reduced to 5.0 to 9.4 post-simplification
Tier 2: Emerging Promise	RAG-grounded guideline decision support; psychosocial support chatbots; tumor board assistance; clinical trial matching	Lammert et al. [76]: 94.7% concordance with expert tumor board; Carl et al. [75]: mean accuracy 76.2%; Akdogan et al. [92]: first RCT showing reduced HADS anxiety scores; Jin et al. [74]: LLM-assisted trial matching
Tier 3: Aspirational /Requires Development	Autonomous adaptive clinical reasoning; real-time intraoperative guidance; validated multilingual and culturally adapted deployment; pan-omics multimodal integration without human review	Kresevic et al. [106]: LLMs underperform physicians in realistic clinical scenarios; Hashimoto et al. [108]: intraoperative AI is highest-value but most technically challenging; Zhu et al. [95]: no general-purpose LLM prospectively validated in agglutinative-language oncology environments

Abbreviations: RCT = randomized controlled trial; MTB = molecular tumor board; WSI = whole-slide imaging; NGS = next-generation sequencing; SaMD = software as a medical device; MLLM = multimodal large language model; HADS = Hospital Anxiety and Depression Scale; RAG = retrieval-augmented generation.

Table 1. Three-Tier Framework of LLM Clinical Readiness in Oncology

Tier	Clinical Domain	Key Evidence
Tier 1: Established Performance	Literature synthesis and evidence mining; clinical documentation; radiology and pathology report simplification; patient education and plain-language communication	Van Veen et al. [40]: 10-fold time reduction in summarization; Ferber et al. [32]: RAG-GPT-4 accuracy 57% to 84%; Kwok et al. [49]: reading level 9.6 to 14 reduced to 5.0 to 9.4 post-simplification
Tier 2: Emerging Promise	RAG-grounded guideline decision support; psychosocial support chatbots; tumor board assistance; clinical trial matching	Lammert et al. [76]: 94.7% concordance with expert tumor board; Carl et al. [75]: mean accuracy 76.2%; Akdogan et al. [92]: first RCT showing reduced HADS anxiety scores; Jin et al. [74]: LLM-assisted trial matching
Tier 3: Aspirational /Requires Development	Autonomous adaptive clinical reasoning; real-time intraoperative guidance; validated multilingual and culturally adapted deployment; pan-omics multimodal integration without human review	Kresevic et al. [106]: LLMs underperform physicians in realistic clinical scenarios; Hashimoto et al. [108]: intraoperative AI is highest-value but most technically challenging; Zhu et al. [95]: no general-purpose LLM prospectively validated in agglutinative-language oncology environments

Abbreviations: RCT = randomized controlled trial; MTB = molecular tumor board; WSI = whole-slide imaging; NGS = next-generation sequencing; SaMD = software as a medical device; MLLM = multimodal large language model; HADS = Hospital Anxiety and Depression Scale; RAG = retrieval-augmented generation.

2. The Paradigm Shift: From Narrow Discrimination to Generative Clinical Reasoning

To appreciate the clinical significance of the post-ChatGPT period, it is instructive to articulate precisely what has changed at the level of architecture and what implications that change carries for oncological practice. The discriminative AI systems that entered clinical oncology in the 2010s were, without exception, task-specific. A convolutional neural network trained to detect lung nodules on CT could not simultaneously interpret a rising carcinoembryonic antigen level, propose a differential diagnosis for a new hepatic lesion, or synthesize a referral letter. Each additional clinical function required a separately trained and validated model, and coordinating multiple narrow models in a coherent clinical workflow presented substantial integration challenges [14]. Their aggregate clinical value, while genuine, could not approach the integrative reasoning that characterizes specialist clinical judgment.

LLMs overcome this constraint through emergent generalization. Singhal and colleagues, reporting the development of Med-PaLM in Nature, demonstrated that a large language model trained on clinical knowledge achieved performance on standardized medical examination questions at the level of a passing physician, a benchmark no narrow discriminative system had approached [15]. GPT-4 was subsequently demonstrated to pass the United States Medical Licensing Examination, as reported by Kung and colleagues in PLOS Digital Health, and to answer subspecialty board questions across multiple disciplines and produce clinical documentation of quality comparable to physician-generated text, as demonstrated by Nori and colleagues (16, 17). These achievements were not demonstrations of clinical deployment readiness, but they established that general-purpose language models possess domain knowledge of sufficient breadth to operate across the clinical spectrum.

Three architectural refinements merit particular attention in the oncological context. The first is RAG, examined in technical detail in Section 3. The second is multimodality: MLLMs capable of processing text, medical images, structured data, and molecular profiles simultaneously, as described by Nam and colleagues and characterized in the survey of Yin and colleagues, represent a qualitative expansion of what AI can contribute to oncological reasoning (18, 19). The third is context extension: contemporary LLMs can process context windows of several hundred thousand tokens, enabling the ingestion and simultaneous reasoning across entire patient records, a capability reviewed by Thirunavukarasu and colleagues in Nature Medicine [20]. These refinements have collectively produced a system capable of operating not merely on isolated data streams but on the integrated clinical picture. The historical transition from task-specific discriminative AI to the current era of generative multimodal LLMs, highlighting the shift toward unified clinical intelligence, is illustrated in Figure 1.

3. Retrieval-Augmented Generation: Technical Architecture, Factual Accuracy, and Grounding

3.1. Static Parametric Knowledge and Its Clinical Limitations

Every LLM trained through standard pre-training procedures acquires what may be termed parametric or static knowledge: a probabilistic encoding of information contained in the training corpus, crystallized at the moment training concluded. As Tu and colleagues described in their review of generalist biomedical AI in NEJM AI, the model's responses are generated by sampling from learned probability distributions over token sequences conditioned on the input context, a process that reflects statistical regularities in the training data rather than access to verified, current, or locally applicable clinical truth [21]. In oncology, where NCCN guidelines may be updated multiple times annually and where new targeted agent approvals, companion diagnostic expansions, and combination therapy data emerge continuously, a model trained on data from a fixed cutoff date carries systematic knowledge gaps that cannot be resolved by prompting alone [22]. A clinician querying an unaugmented LLM about frontline treatment selection in a rapidly evolving indication receives a response conditioned on the statistical distribution of the training corpus rather than on the most recent evidence, without any mechanism to distinguish one from the other. The fundamental distinction between static parametric knowledge and RAG-grounded dynamic retrieval is summarized in Table 2.

3.2. Probabilistic Text Generation and Factual Inaccuracy in Clinical Outputs

A fundamental characteristic of autoregressive language models is that token generation is governed by learned conditional probability distributions, not by reference to a verified external ground truth. Ji and colleagues, in a comprehensive survey of hallucination in natural language generation published in ACM Computing Surveys, classified this failure mode into intrinsic forms, in which the model contradicts retrievable source material, and extrinsic forms, in which the model generates information not present in any identifiable source [23]. The extrinsic category is the more clinically dangerous in ungrounded deployments. Alkaissi and McFarlane documented specific instances of medically consequential inaccuracies in ChatGPT outputs in the journal Cureus, including fabricated clinical trial citations with plausible author names and journal titles, invented drug approval dates, and nonexistent dosing protocols [24]. Huang and colleagues, in a technical survey of factual inaccuracy mechanisms in large language models published on arXiv, provided a taxonomy of the mechanisms by which autoregressive generation produces incorrect outputs and identified grounding-based mitigation strategies as the most effective class of solutions [25]. Omar and colleagues demonstrated in Communications Medicine that six leading LLMs, when presented with clinical vignettes containing a single planted false data point, reproduced or elaborated on the error in up to 83% of cases across trials [26]. Gundabathula and colleagues, developing a framework for assessing clinical safety and factual accuracy rates in LLMs for medical text summarization published in npj Digital Medicine, observed that structured and constrained generation tasks exhibited substantially lower error rates at 1.47% compared to open-ended clinical reasoning tasks [27]. Moell and Aronsson synthesized these findings into a structured harm reduction framework for medical LLM use in the Journal of Medical Internet Research, identifying high-stakes clinical specialties including oncology as requiring the most conservative deployment configurations and emphasizing mandatory clinician verification as the non-negotiable safety layer [28].

Table 3. Hallucination Taxonomy and Evidence-Based Mitigation Strategies for Clinical Oncology LLM Deployment

Hallucination Type	Definition	Clinical Example in Oncology	Evidence-Based Mitigation Strategy
Intrinsic hallucination	Model output contradicts retrievable source material in its own training corpus	LLM states incorrect first-line dosing for a standard regimen documented in its own training data	RAG grounding with authoritative guideline retrieval; chain-of-thought prompting [35]
Extrinsic hallucination	Model generates information not present in any identifiable source, which is the more clinically dangerous category	Fabricated clinical trial citations with plausible author names; invented drug approval dates; nonexistent dosing protocols [24]	RAG constrains generation to retrieved verified documents; claim-level fact-checking against structured medical databases [34]
Adversarial hallucination	Model reproduces or elaborates on planted false information from the prompt	LLM elaborates on incorrect patient history data in 83% of vignettes containing a planted error [26]	Structured and constrained generation tasks exhibiting 1.47% error rate versus open-ended tasks [27]; mandatory clinician verification [28]
Knowledge cutoff gap	Outdated recommendations due to training data cutoff in rapidly evolving indications	Recommendation of a superseded first-line agent following new trial data published after training cutoff	RAG with continuously updated index, given that NCCN guidelines may be updated multiple times annually [22]
Black-box opacity	No audit trail identifies which training data or weights produced a specific erroneous output	Incorrect treatment recommendation with no identifiable causal chain; complicates attribution of liability	Explainable AI architectures; verifiable cited responses as in ASCO Guidelines Assistant model [37]; human-in-the-loop as primary accountability mechanism [80]

Abbreviations: RAG = retrieval-augmented generation; NCCN = National Comprehensive Cancer Network; LLM = large language model; CoT = chain-of-thought prompting.

Table 3. Hallucination Taxonomy and Evidence-Based Mitigation Strategies for Clinical Oncology LLM Deployment

Hallucination Type	Definition	Clinical Example in Oncology	Evidence-Based Mitigation Strategy
Intrinsic hallucination	Model output contradicts retrievable source material in its own training corpus	LLM states incorrect first-line dosing for a standard regimen documented in its own training data	RAG grounding with authoritative guideline retrieval; chain-of-thought prompting [35]
Extrinsic hallucination	Model generates information not present in any identifiable source, which is the more clinically dangerous category	Fabricated clinical trial citations with plausible author names; invented drug approval dates; nonexistent dosing protocols [24]	RAG constrains generation to retrieved verified documents; claim-level fact-checking against structured medical databases [34]
Adversarial hallucination	Model reproduces or elaborates on planted false information from the prompt	LLM elaborates on incorrect patient history data in 83% of vignettes containing a planted error [26]	Structured and constrained generation tasks exhibiting 1.47% error rate versus open-ended tasks [27]; mandatory clinician verification [28]
Knowledge cutoff gap	Outdated recommendations due to training data cutoff in rapidly evolving indications	Recommendation of a superseded first-line agent following new trial data published after training cutoff	RAG with continuously updated index, given that NCCN guidelines may be updated multiple times annually [22]
Black-box opacity	No audit trail identifies which training data or weights produced a specific erroneous output	Incorrect treatment recommendation with no identifiable causal chain; complicates attribution of liability	Explainable AI architectures; verifiable cited responses as in ASCO Guidelines Assistant model [37]; human-in-the-loop as primary accountability mechanism [80]

Abbreviations: RAG = retrieval-augmented generation; NCCN = National Comprehensive Cancer Network; LLM = large language model; CoT = chain-of-thought prompting.

3.3. RAG as Architectural Grounding: From Unconstrained Sampling to Verifiable Retrieval

RAG was introduced by Lewis and colleagues as a framework for knowledge-intensive natural language processing tasks, published in Advances in Neural Information Processing Systems, and addresses the probabilistic accuracy problem at the architectural rather than the prompting level [29]. In oncology, where a hallucinated drug dose or a fabricated contraindication is not a statistical nuisance but a direct patient safety event, this architectural distinction is the difference between a tool that can and cannot be responsibly deployed. RAG transforms the LLM from a conversational chatbot drawing on memorized probabilities into a real-time guideline auditor: a system that scans NCCN and ASCO corpora in seconds, retrieves the specific passage authorizing or qualifying a clinical recommendation, and presents that passage as the verifiable foundation of its response. In a RAG architecture, an LLM is connected at inference time to an external retrieval system, typically a dense vector index built over a curated document corpus. When a query is submitted, the retrieval component performs a semantic similarity search over the index, returns the most relevant document passages as context, and presents these passages to the LLM alongside the original query. The LLM is instructed to generate its response using only the retrieved material, converting the generation process from unconstrained parametric sampling to constrained contextual reasoning over retrieved, inspectable source documents. This architectural transformation converts the system from a closed-book examiner relying on memorized probabilities to an open-book reasoner operating over specified, verifiable, and continuously updatable evidence. Zarfati and colleagues, reviewing RAG applications in oncology in the European Journal of Cancer, concluded that RAG-grounded systems consistently outperformed non-RAG counterparts across accuracy, citation quality, and clinical safety metrics, and proposed RAG grounding in authoritative oncological guidelines as a prerequisite rather than an enhancement for clinical deployment [30]. Xiong and colleagues demonstrated in the Pacific Symposium on Biocomputing that iterative follow-up question strategies in RAG-based medical systems further reduced residual inaccuracy rates by improving retrieval recall for complex multi-part queries [31]. The operational workflow of Retrieval-Augmented Generation (RAG), which anchors LLM outputs in authoritative oncology guidelines like NCCN to minimize hallucinations, is depicted in Figure 2.

3.4. Empirical Evidence for RAG-Grounded Accuracy in Oncology

Empirical evidence for the accuracy-improving effects of RAG in oncology is now substantial. Ferber and colleagues, publishing in NEJM AI, demonstrated that RAG-augmented GPT-4 reduced incorrect response rates from 43% to 16% in oncology guideline retrieval and comparison tasks, while simultaneously improving overall accuracy from 57% to 84% [32]. Liu and colleagues developed a RAG pipeline grounded in NCCN guidelines and a standard radiation oncology reference text, reporting in a preprint on medRxiv that the system achieved 91.5% accuracy on American College of Radiology Training Examination questions, substantially exceeding ungrounded model performance [33]. Zeba and colleagues developed a domain-specific healthcare LLM framework combining model fine-tuning with an LLM-independent granular fact-checking module, demonstrating improved factual reliability in clinical text summarization [34]. Wei and colleagues, in a paper on chain-of-thought prompting published in Advances in Neural Information Processing Systems, demonstrated that instructing the model to reason step-by-step before arriving at a final response reduced error rates in complex multi-step reasoning tasks by making the logical path transparent and interruptible [35].

3.5. Fine-Tuning Versus RAG: Distinct Mechanisms and Complementary Roles

Domain-specific fine-tuning, in which a model's parameters are updated through continued training on a curated clinical corpus, represents a complementary but mechanistically distinct approach. The Meditron Team, reporting the development of Meditron-70B on arXiv, developed a Llama-based model fine-tuned on medical literature including PubMed abstracts, clinical guidelines, and medical examination questions that demonstrates improved performance on medical knowledge benchmarks relative to the base model, but whose knowledge currency remains bounded by its training corpus [36]. The practical implication is that fine-tuning and RAG address different dimensions of the clinical LLM performance problem and are best understood as complementary components of a comprehensive deployment architecture. The ASCO Guidelines Assistant, developed in collaboration with Google Cloud using the Vertex AI platform and Gemini models, as reported by Cavallo in The ASCO Post, combines a foundation model with strong medical competence with RAG grounding in ASCO’s current guideline documents, providing highlighted citations linking to the specific source sentence for every answer generated [37]. This design instantiates the grounded, transparent, and human-in-the-loop architecture identified by the EU AI Act of 2024 as appropriate for high-risk clinical AI deployment [38].

4. Domains of Established Performance: Evidence-Supported Clinical Utility

4.1. Literature Synthesis, Evidence Mining, and Clinician Time Recovery

The biomedical literature grows at a pace that renders comprehensive monitoring impossible for any individual clinician. PubMed indexes in excess of one million new records annually, and the oncology literature expands at a rate that consistently outpaces the reading capacity of subspecialists [39]. Van Veen and colleagues, publishing in Nature Medicine, demonstrated that LLMs fine-tuned for clinical summarization outperformed general-purpose medical experts in producing accurate and comprehensive summaries of complex clinical texts, with an estimated ten-fold reduction in summarization time relative to trained clinician reviewers [40]. Gates and colleagues, publishing in Systematic Reviews, demonstrated that technology-assisted title and abstract screening reduces the screening phase by 40% to 70% while maintaining sensitivity for relevant study identification comparable to dual human review [41]. Truhn and colleagues, publishing in the Journal of Pathology, demonstrated that GPT-4 extracted structured oncological data from unstructured histopathology reports at accuracy competitive with manual curation while processing thousands of reports within the time a trained human annotator would require to process dozens [42]. Benary and colleagues, publishing in JAMA Network Open, demonstrated LLM application to molecular tumor board documentation, reporting a 60% reduction in documentation time with accuracy maintained comparable to unaided clinician documentation [43]. Sun and colleagues, publishing in the Journal of Clinical Oncology, demonstrated that LLMs extracted immune-related adverse event data from unstructured clinical notes with sensitivity exceeding International Classification of Diseases coding [44]. Evidence supporting established performance across key application domains is consolidated in Table 4.

4.2. Administrative Burden Reduction and Clinical Documentation Quality

Physician burnout in oncology has reached levels recognized as a systemic threat, with documentation burden consistently identified as a primary contributor, as reviewed by West, Dyrbye, and Shanafelt in the Journal of Internal Medicine [45]. Umeton and colleagues, reporting the first large-scale institute-wide deployment of GPT-4 in a cancer center in NEJM AI, documented quantified time savings in documentation tasks across multiple clinical departments and identified the reduction in iterative correction cycles as the primary source of efficiency gain [46]. Busch and colleagues, publishing a systematic review of LLM applications in patient care in Communications Medicine, confirmed across a large synthesized evidence base that LLM-generated clinical documentation drafts reduce total documentation burden even when they require clinician editing before finalization [12]. Moradinasab and colleagues, publishing in JCO Clinical Cancer Informatics, reviewed LLM applications specifically in population oncology and confirmed administrative efficiency gains as the most consistently replicated finding across the literature [13].

4.3. Patient Education, Health Literacy, and Plain-Language Communication

Health literacy, defined as the capacity of individuals to obtain, process, and understand basic health information needed to make appropriate health decisions, is unevenly distributed and systematically inadequate for the complexity of oncological information, as established by Berkman and colleagues in Annals of Internal Medicine [47]. Al Hussein Al Awamlh and colleagues, publishing in Cancer, demonstrated an independent association between health literacy and all-cause mortality in cancer patients, establishing a direct clinical rationale for interventions that reduce the literacy burden of oncological information [48]. Kwok and colleagues, publishing a systematic review of ChatGPT-based radiology report simplification in the Journal of Medical Imaging and Radiation Oncology, found median reading grade levels of original reports ranging from 9.6 to 14, with post-simplification outputs ranging from 5.0 to 9.4, factual accuracy from 78% to 100%, and report completeness from 83% to 100% [49]. Yang and colleagues, publishing in the Journal of Medical Internet Research, analyzed 698 malignant tumor radiology reports from Chinese oncology centers and demonstrated that GPT-4-generated simplified interpretive reports improved correct classification of report outcomes from 63.9% to 83.3% with a mean patient satisfaction score of 8.30 out of 10 [50]. Prucker and colleagues, publishing in Lancet Digital Health, confirmed across a larger synthesized evidence base that LLM-based report simplification improves comprehension and achieves high patient satisfaction with manageable error rates when appropriate clinician review protocols are maintained [51]. Chung and colleagues, publishing in Digital Health, demonstrated that ChatGPT achieved sixth-grade reading level targets for a substantial proportion of complex MRI reports evaluated against the American Medical Association recommendation [52]. Haver and colleagues, publishing in Radiology Imaging Cancer, confirmed comparable simplification efficacy specifically in breast cancer patient-facing materials [53]. Steimetz and colleagues, publishing in JAMA Network Open, demonstrated utility in pathology report simplification with accuracy acceptable for clinical use [54]. Lyu and colleagues, publishing in Visual Computing for Industry, Biomedicine, and Art, demonstrated that structured prompt engineering substantially improved plain-language translation accuracy for radiology reports [55]. Doshi and colleagues, publishing in Radiology, provided quantitative evaluation of LLMs for streamlining radiology report impressions, confirming meaningful accuracy improvements with structured prompting approaches [56].

4.4. Conversational Patient Support and Chatbot Applications

LLM-powered chatbots represent a qualitatively different modality of patient information delivery, accommodating personalized, iterative, bidirectional exchange. Ayers and colleagues, publishing in JAMA Internal Medicine, found that AI responses to patient medical questions were rated significantly higher on quality and empathy dimensions relative to physician responses, generating substantive discussion about the communicative dimensions of clinical care currently underserved by time-pressured clinical encounters [57]. Sorin and colleagues, publishing in npj Breast Cancer, evaluated ChatGPT as a support tool for breast tumor board discussions and reported that oncologists rated LLM-generated case summaries and guideline references as useful in 78% of evaluations [58]. Zhou and colleagues, publishing in the International Journal of Surgery, compared LLM chatbot performance against oncology physicians on colorectal cancer queries, finding performance comparable to junior physicians on informational tasks and inferior to senior specialists on complex diagnostic reasoning [59]. Lopez and colleagues, publishing a systematic review in Surgical Oncology, comprehensively reviewed ChatGPT applications in colorectal cancer and confirmed the pattern of informational utility with diagnostic reasoning limitations [60]. Nov and colleagues, reporting on medRxiv, documented average patient trust scores for chatbot medical advice of approximately 3.4 out of 5, with trust calibrated lower for more complex medical questions [61].

5. Multimodal Large Language Models and the Pan-Omics Vision in Precision Oncology

5.1. The Multimodal Imperative: Beyond Text and Beyond Radiology

The concept of MLLMs warrants dedicated and technically detailed attention because it addresses the most fundamental characteristic of oncological clinical data: its profound multimodality. As Acosta, Falcone, Rajpurkar, and colleagues established in a comprehensive review of multimodal biomedical AI in Nature Medicine, a complete contemporary oncological assessment integrates CT, MRI, PET, and ultrasound imaging; macroscopic and microscopic pathological analysis including WSI interpretation; molecular genomic profiling via NGS; transcriptomic and proteomic characterization; liquid biopsy results including circulating tumor DNA; radiomics feature extraction from volumetric imaging; RWE from clinical registries; patient-reported outcome measures; and sociodemographic determinants of health [62]. No single data modality is sufficient for management decisions in most cancer types, and the clinical value of oncology subspecialty practice resides precisely in the integration of these streams into a coherent, individualized decision framework. As Yin and colleagues established in their survey of multimodal large language models in the National Science Review, and as Nam and colleagues described in their review in the Korean Journal of Radiology, the aspiration of MLLMs is to enable AI systems to perform this integration with computational consistency and encyclopedic evidence coverage (18, 19).

It is important to distinguish the MLLM vision from the radiology report simplification applications reviewed in Section 4. Report simplification is a text-to-text task in which an LLM processes a natural language radiology report and generates a simplified version. True multimodality requires a model capable of simultaneously processing and reasoning across raw imaging data, structured molecular assay outputs, natural language clinical notes, and quantitative laboratory results, and generating a synthesized clinical inference that could not be produced from any single input stream alone [62]. MLLM performance across key oncology imaging and pathology tasks is consolidated in Table 5.

5.2. Radiology: From Report Simplification to Active Diagnostic Reasoning

In diagnostic radiology, MLLM performance has advanced rapidly across successive model generations. Suh and colleagues, publishing in Radiology, demonstrated that GPT-4V achieved top-3 differential diagnosis accuracy in the Diagnosis Please educational series that was not statistically different from subspecialty-trained board-certified radiologists at approximately 49% [63]. Bousaid and colleagues, publishing in European Radiology, evaluated GPT-4V multimodal performance across 230 consecutive emergency radiology imaging studies and found correct imaging modality identification in 100% of cases, correct anatomical region identification in 87.1%, and pathology identification accuracy at 35.2%, illustrating both the progress achieved and the distance remaining from autonomous image interpretation [64]. Hou and colleagues, publishing in Radiology, demonstrated that OpenAI's o1 and GPT-4o achieved accuracy rates of 59% and 54% respectively on RSNA 2024 Case of the Day examination questions [65]. Han and colleagues, publishing in Academic Radiology, compared GPT-4 and Gemini across 205 consecutive patients with paired serial abdominal CT reports and found that GPT-4 achieved 96.2% accuracy in matching oncologically relevant findings between serial reports, with superior identification of tumor status change compared to Gemini [66]. Nakaura and colleagues, publishing in the Japanese Journal of Radiology, provided a comprehensive guide for radiologists on these developments, characterizing the trajectory from report simplification toward integrated diagnostic assistance [67]. Doshi and colleagues, publishing in Radiology, provided quantitative evaluation confirming meaningful accuracy improvements with structured prompting in radiology report streamlining [56].

5.3. Computational Pathology and Whole-Slide Image Analysis

Computational pathology has historically demanded large, task-specific, expert-annotated training datasets for each histological classification task. Kather and colleagues, publishing in Nature Communications, demonstrated that GPT-4V with in-context learning matched or outperformed specialized neural networks on colorectal cancer tissue subtype classification, colon polyp subtyping, and breast tumor detection in lymph nodes while requiring minimal labeled training data, establishing proof of concept for general-purpose MLLM application to computational pathology [68]. Chen and colleagues, publishing in Nature Medicine, developed CONCH, a general-purpose foundation model for computational pathology demonstrating competitive performance on tissue classification, disease subtyping, and survival prediction tasks across multiple cancer types in a self-supervised learning paradigm [69]. Lu and colleagues, publishing in Nature Medicine, introduced a visual-language foundation model trained on paired pathology image-text data, achieving state-of-the-art performance on WSI classification and zero-shot diagnostic tasks [70]. These models represent a category of clinical AI tool that can be adapted to specific oncological tasks with minimal additional annotation, with direct implications for deployment in resource-limited settings where large annotated datasets are unavailable.

5.4. Radiomics Integration: Quantitative Imaging Phenotyping and Language-Based Reasoning

Radiomics, the high-throughput extraction of quantitative features from standard-of-care imaging studies, has established that volumetric CT, MRI, and PET data encodes prognostic and predictive information substantially beyond what is accessible to visual inspection. Lambin and colleagues, publishing in Nature Reviews Clinical Oncology, demonstrated that radiomic feature sets extracted from primary tumor volumes can predict pathological response to neoadjuvant chemotherapy, regional lymph node involvement, molecular subtype assignment, and overall survival across multiple cancer types in independent validation cohorts [71]. MLLMs offer a fundamentally different integration architecture in which a model with appropriate multimodal encoders can receive radiomic, genomic, and clinical data types as input within a single context window and reason across them using the model's pre-existing clinical knowledge, mirroring the multidisciplinary tumor board process. Boehm and colleagues, in their review of multimodal data integration in precision oncology published in Nature Reviews Cancer, characterized the convergence of imaging, pathology, genomics, transcriptomics, and clinical data as the defining technical challenge of next-generation oncological AI and identified the MLLM architecture as the most promising framework for addressing it [72].

5.5. Pan-Omics Integration: NGS, Transcriptomics, and Proteomics in Precision Oncology

The precision oncology paradigm rests on the premise that therapeutic decision-making should be guided by the molecular characteristics of the individual tumor. The realization of this premise at clinical scale requires integration of NGS data encoding somatic mutation profiles, copy number alterations, and structural variants; RNA sequencing data characterizing tumor transcriptional programs and immune microenvironment composition; proteomic data reflecting post-translational modifications and pathway activation states; and epigenomic data encoding chromatin accessibility and DNA methylation patterns. Chen and colleagues, publishing in Cancer Cell, demonstrated that multimodal foundation models trained on paired imaging and genomic data from The Cancer Genome Atlas could predict molecular subtypes from imaging features alone and identify imaging correlates of targetable mutations, establishing a proof of concept for imaging-genomic integration within a single model architecture [73]. Boehm and colleagues, in Nature Reviews Cancer, characterized this convergence as the defining technical challenge of next-generation oncological AI [72]. Jin and colleagues demonstrated that an LLM-based framework could accurately match patients to clinical trials, achieving 87.3% accuracy in criterion-level eligibility assessment and reducing expert screening time by 42.6%, highlighting the potential of LLM-assisted trial matching in precision oncology [74]. The convergence of radiology, computational pathology, radiomics, NGS, transcriptomics, proteomics, and RWE within a single MLLM reasoning framework constitutes the technical vision of AI-integrated precision oncology. Current demonstrations are proofs of concept rather than clinical deployments, but the architectural feasibility of the integrated vision is now established (62, 72). The integration of heterogeneous data streams, including genomics, radiology, and pathology, into a centralized multimodal AI node for precision oncology is visualized in Figure 3.

6. Domains of Emerging Promise: Guideline-Grounded Decision Support and Psychosocial Applications

6.1. The ASCO Guidelines Assistant and the Institutional Authority Axis

The integration of LLMs into clinical decision support represents the most consequential application space in contemporary oncology AI. ASCO's launch of the ASCO Guidelines Assistant, announced in May 2025 and reported by Cavallo in The ASCO Post in September 2025, is the most significant institutional endorsement to date. The system, developed with Google Cloud's Vertex AI platform and Gemini models, constrains all responses to ASCO's own evidence-based guideline documents and accompanies every response with a highlighted citation linking to the specific source sentence [37]. The system is designed to report absence of guideline coverage rather than to extrapolate. This architecture, described by ASCO's Chief Executive Officer Clifford Hudis as designed to "quickly surface the answer to a question in a way that establishes reliability and trust," instantiates the RAG-grounded, verifiable approach identified in the clinical literature as the prerequisite for safe LLM deployment in high-stakes medical settings.

Carl and colleagues, in their systematic review and meta-analysis in npj Precision Oncology covering 34 studies, reported average overall accuracy of 76.2% for LLMs in clinical oncology applications, with diagnostic accuracy at 67.4% [75]. Lammert and colleagues, publishing in JCO Precision Oncology, demonstrated 94.7% concordance with expert tumor board recommendations for their RAG-based MEREDITH system in precision oncology applications, and identified treatment options not initially considered by the human board [76]. Rydzewski and colleagues, publishing the NEJM AI comparative evaluation across multiple LLMs on clinical oncology questions across subspecialties, documented significant performance variance attributable to model architecture, prompt engineering strategy, and oncological subdomain, underscoring the importance of domain-specific evaluation prior to deployment [77]. Iannantuono and colleagues, publishing in The Oncologist, reported higher rates of fully correct responses with GPT-4 (75.4%) than with GPT-3.5 (58.5%) and Google Bard (43.8%) across immuno-oncology questions [78].

National Comprehensive Cancer Network guidelines, as the most widely consulted clinical practice guidelines in oncology, provide a complementary anchor for RAG-based systems [79]. Liu and colleagues demonstrated that a RAG pipeline grounded specifically in NCCN guidelines achieved 91.5% accuracy on standardized radiation oncology examination questions [33]. These findings collectively support the proposition that RAG grounding in NCCN and ASCO guidelines is not merely a technical enhancement but the architectural prerequisite for a system that can function as a reliable clinical decision support tool, not an autonomous decision-maker. The system is designed to report absence of guideline coverage rather than to extrapolate. Benchmarking evidence across multiple LLMs and study designs is consolidated in Table 6.

6.2. Human-in-the-Loop: Clinical, Legal, and Regulatory Positioning of LLMs as Decision Support Tools

The human-in-the-loop concept, as applied to clinical AI governance and described by Sittig and Singh in their safety recommendations published in JAMA, specifies that AI systems involved in consequential clinical decisions must be designed to require review, validation, and authorization by a qualified human clinician before any AI-generated recommendation is acted upon [80]. In oncology, where chemotherapy agents and targeted therapies represent some of the highest-risk pharmacological interventions in clinical practice, this requirement is a substantive patient safety constraint, not a regulatory formality.

The FDA's framework for AI and machine learning as SaMD, published in 2024, distinguishes between locked and adaptive algorithms. LLMs as continuously updated foundation models fall into the adaptive category and are subject to the most stringent oversight requirements, including pre-market review for high-risk clinical applications and post-market performance monitoring [81]. The EU AI Act of 2024, enacted by the European Parliament, classifies medical decision-support AI as high-risk and imposes requirements for transparency, human oversight, accuracy documentation, and post-market surveillance that create a de facto human-in-the-loop mandate for any LLM deployed in a clinical decision context [38]. Shumway and Hartman, publishing in the Journal of Osteopathic Medicine, concluded in their legal analysis that the current U.S. legal framework is inadequately prepared for AI-assisted clinical decision-making and proposed a tiered liability model in which responsibility is allocated proportionally among developers, institutional deployers, and attending clinicians according to their degree of control over the AI system's configuration and use [82]. ESMO's ELCAP guidance, published by Verlingue and colleagues in Lancet Regional Health Europe in 2025, articulates principles of transparent AI limitations disclosure, mandatory human oversight, and systematic outcome evaluation representing the oncology community's professional consensus on governance standards [83]. The practical implication is that LLM outputs must be documented as decision-support inputs subject to independent clinician review and authorization, not as autonomous clinical decisions, in both system design and clinical workflow documentation.

6.3. Benchmarking Performance Across Oncology Subspecialties

Longwell and colleagues, publishing in JAMA Network Open, evaluated LLM performance on medical oncology board examination questions and found GPT-4 at passing performance with significant variability by question type, with factual recall questions consistently outperforming complex clinical reasoning tasks [84]. Thirunavukarasu and colleagues, in a comprehensive review in Nature Medicine, established a taxonomy of LLM applications in medicine and identified the gap between knowledge retrieval performance and adaptive clinical reasoning as the central challenge for medical AI evaluation [20]. Cabral and colleagues, applying Script Concordance Testing to multiple LLMs and publishing in JAMA Internal Medicine, found performance substantially lower than on multiple-choice knowledge tests, revealing that high examination scores reflect memorized knowledge rather than adaptive reasoning under conditions of uncertainty [86]. Bedi and colleagues, publishing a systematic review of 519 studies in JAMA, confirmed that the vast majority of published LLM healthcare evaluations measure technical accuracy rather than patient-centered clinical outcomes, identifying prospective clinical trials with validated endpoints as the urgent priority for the field [87]. Singhal and colleagues, reporting the extension of Med-PaLM evaluation toward expert-level medical question answering on arXiv, demonstrated continued improvement through instruction tuning and reinforcement learning from human feedback [88]. Saab and colleagues, reporting Gemini model capabilities in medicine on arXiv, characterized model performance across multiple specialties identifying specific subdomain strengths and weaknesses [89]. Yang and colleagues, reviewing LLM development, applications, and challenges in healthcare in Health Care Science, provided a comprehensive comparative analysis of model architectures and their clinical performance characteristics [90].

6.4. Psychosocial Support and Mental Health Adjuncts in Cancer Care

Mitchell and colleagues established in Lancet Oncology that anxiety and depression are prevalent in newly diagnosed cancer patients at rates of 20% to 40%, yet access to formal psycho-oncological support is constrained in most health systems by workforce shortages, cost, and geographic inaccessibility [91]. Akdogan and colleagues, conducting a prospective randomized controlled trial in 150 chemotherapy-naive cancer patients published in the European Journal of Cancer, demonstrated that a ChatGPT-based digital counseling intervention produced significantly lower Hospital Anxiety and Depression Scale scores prior to the second chemotherapy cycle compared to standard care, constituting the first randomized evidence of LLM-based psychological intervention efficacy in cancer patients [92]. Zhang and colleagues, publishing in Frontiers in Digital Health, demonstrated feasibility and preliminary efficacy of generative AI chatbots for psychological burden reduction in pediatric and adolescent cancer patients [93]. Farzan and colleagues, publishing a systematic review of 160 studies in npj Mental Health Research, confirmed that LLM-based chatbots constituted 45% of new studies published in 2024 but that only 16% had undergone clinical efficacy testing [94]. Rigorous randomized controlled trials with validated psychometric endpoints and adequate follow-up periods are identified as an urgent research priority. The evidence base for LLM-based psychosocial applications in oncology is consolidated in Table 7.

7. Global Versus Local Challenges: Linguistic Barriers, Cultural Bias, and Non-English Clinical Contexts

7.1. English-Language Training Data Dominance and Multilingual Performance Limitations

The foundational LLMs that underpin clinical applications were trained predominantly on English-language text, reflecting the language composition of the majority of biomedical literature, clinical documentation systems, and publicly available health information repositories in the training corpora. Zhu and colleagues, publishing a systematic review of multilingual large language models in medicine in npj Digital Medicine, estimated English-language content in the training data of frontier LLMs at 70% to 92%, with the remainder distributed among dozens of other languages at substantially lower representation, and demonstrated that models exhibiting strong performance on English-language medical tasks show substantially degraded accuracy and increased error rates when operating in languages with lower training corpus representation [95]. Omiye, Gui, Rezaei, and colleagues, publishing in Annals of Internal Medicine a review of the potentials and pitfalls of large language models in medicine, confirmed that performance degradation in non-English clinical contexts is a reproducible and clinically significant finding across multiple evaluation frameworks [96].

Morphologically complex agglutinative languages with smaller online medical corpora than English represent a specific and underexamined challenge for clinical LLM deployment. Medical terminology in such languages typically encompasses both Latinized international terminology and vernacular equivalents that may differ across institutional, generational, and regional contexts. Turkish serves as an instructive example of this broader class. Yildirim and colleagues, publishing in Annals of Clinical and Analytical Medicine, demonstrated that while GPT-4 achieved acceptable simplification accuracy for straightforward Turkish radiology reports, performance degraded measurably on reports containing institution-specific terminology, local abbreviations for treatment regimens, and references to regionally prevalent cancer subtypes with distinct clinical nomenclature [97]. Comparable degradation patterns have been documented in Arabic, Japanese, and other non-Western clinical languages. These findings collectively illustrate that performance characteristics reported in English-language benchmarking studies cannot be assumed to generalize to other clinical deployment environments without local validation, and that the systematic absence of such validation data represents a global research gap, not merely a language-specific one. The full scope of global and local deployment challenges is consolidated in Table 8.

7.2. Healthcare System-Specific Knowledge: Reimbursement Frameworks and Formulary Constraints

A dimension of local clinical relevance that is entirely absent from the training corpora of general-purpose LLMs is health system-specific reimbursement and formulary knowledge. In most national healthcare systems, oncology treatment decisions are substantially constrained by formulary management bodies that specify eligible indications, required biomarker testing, prior treatment line requirements, and dosing schedules for reimbursable agents. Germany manages this through the Gemeinsamer Bundesausschuss, the United Kingdom through the National Institute for Health and Care Excellence, and middle-income health systems through analogous national frameworks. An LLM providing treatment recommendations, even if its NCCN-grounded guideline knowledge is current and accurate, may generate recommendations for agents or indications not covered under the applicable reimbursement framework, or may fail to flag the documentation requirements that must be satisfied before approval. This represents a category of local contextual knowledge that requires either dedicated local knowledge base development or integration with reimbursement-specific retrieval indices, and constitutes a universal rather than geographically specific challenge for international LLM deployment (96, 98).

7.3. Population Genetics, Ancestral Variation, and Oncological Implications

Beyond linguistic and administrative considerations, population-specific biological characteristics represent a domain in which general-purpose LLMs trained predominantly on Western medical literature may generate systematically biased recommendations for non-Western patient populations. This challenge is not unique to any single country: allele frequency distributions for cancer-relevant variants differ meaningfully between Southern European, South Asian, East Asian, Middle Eastern, and Sub-Saharan African populations relative to the Northern European reference cohorts that dominate the training literature. As one documented example, Yazici, publishing in Hereditary Cancer in Clinical Practice, reported that BRCA1 and BRCA2 pathogenic variant spectra in hereditary breast-ovarian cancer families from Turkey include founder mutations not prominently represented in databases that inform standard LLM training [99]. Tunca and colleagues, publishing in the International Journal of Colorectal Disease, similarly documented that MLH1, MSH2, and MSH6 variant distributions in Lynch syndrome families from Turkey differ from Northern European reference cohorts [100]. The clinical implication generalizes across diverse populations: an LLM providing hereditary cancer risk assessment or targeted therapy eligibility guidance without awareness of population-specific variant frequencies may underestimate or mischaracterize risk with direct consequences for genetic counseling, preventive intervention, and treatment selection. The systematic underrepresentation of non-Western populations in the genomic databases that inform LLM training is a global equity problem, and correcting it requires investment in population-specific genomic data resources as a prerequisite for equitable AI-assisted oncological care worldwide.

7.4. Cultural Communication Norms and Patient-Facing Application Adaptation

The health literacy dimension of LLM application in oncology is further complicated by cultural factors that extend beyond language. Communication of cancer diagnoses, prognosis, and end-of-life care involves cultural norms regarding family involvement in decision-making, disclosure practices, illness representations, and attitudes toward uncertainty that differ substantially across cultural contexts and that are not encoded in the English-language medical literature that constitutes the primary training corpus of frontier LLMs [95]. Publishing in Supportive Care in Cancer, documented higher rates of family-mediated disclosure practices in the clinical context and preferences for physician authority in treatment decision-making that differ from the autonomous patient decision-making model underlying much of the patient communication literature on which LLM training data draws [101]. An LLM-powered patient communication tool intended for deployment in oncology must incorporate awareness of these cultural preferences in its communication style, disclosure approach, and shared decision-making facilitation if it is to serve patients effectively rather than imposing a culturally dissonant communication framework (97, 101). Genuine cultural adaptation, distinct from mere translation, represents a specific research and development priority for the field. The current landscape of oncology AI research exhibits a significant lack of diversity, as the geographic concentration of clinical validation studies and the resulting representation gaps in global cancer literature are illustrated in Figure 4.

8. Algorithmic Bias, Data Privacy, Legal Liability, and Ethical Dimensions

8.1. Algorithmic Bias and Underrepresented Populations

LLMs are trained on corpora that predominantly reflect the medical literature, clinical records, and health information systems of high-income, largely English-speaking, Western countries. Omiye and colleagues, publishing in npj Digital Medicine, demonstrated that several major LLMs propagate race-based medical misconceptions, including discredited biological associations between racial group and physiological parameters, in ways that could directly influence clinical management [102]. Omar and colleagues, publishing a systematic review in the International Journal of Equity Health, confirmed that performance disparities across patient age, sex, race, and socioeconomic status are reproducible across multiple LLMs and evaluation contexts, and proposed minimum demographic performance reporting standards as a prerequisite for clinical deployment [103]. The clinical implications for oncology in diverse populations are detailed in Section 7. Demographic performance stratification, with stratified accuracy reporting across racial, ethnic, and linguistic subgroups as a prerequisite for clinical deployment authorization, is identified as an urgent methodological standard for the field. Clinical institutions deploying LLMs should actively monitor for outcome disparities across patient demographic groups and establish feedback mechanisms for identifying and correcting performance gaps as they emerge.

8.2. Data Privacy: HIPAA, GDPR, and the KVKK Framework

Data privacy compliance in cloud-based LLM deployments requires jurisdiction-specific legal analysis. In the United States, the Health Insurance Portability and Accountability Act requires business associate agreements with AI vendors, data minimization practices in API calls, and audit trails for AI-assisted clinical decisions. In Europe, the General Data Protection Regulation (GDPR) imposes explicit legal basis requirements, data subject rights to explanation and contestation, and restrictions on extra-EEA data transfers [104]. Most middle-income and upper-middle-income countries have enacted analogous national data protection statutes in recent years, creating a patchwork of jurisdiction-specific compliance obligations for any organization deploying cloud-based LLMs internationally. Locally deployed open-source LLMs based on the Llama architecture address data sovereignty concerns at the cost of performance reduction; Meditron-70B represents the most mature medically fine-tuned open-source alternative [36]. Putzier and colleagues, publishing in npj Digital Medicine, analyzed cloud computing implementation in the German healthcare system under GDPR and identified contractual frameworks, technical pseudonymization, and institutional data governance as prerequisites for compliant deployment, conclusions applicable across international deployment contexts [105].

8.3. Legal Liability and the SaMD Regulatory Pathway

The attribution of legal responsibility for AI-assisted clinical errors remains unresolved in most jurisdictions. Shumway and Hartman, publishing in the Journal of Osteopathic Medicine, proposed a tiered liability model allocating responsibility proportionally among developers, institutional deployers, and attending clinicians [82]. This legal ambiguity is compounded by the opacity inherent in large language model architectures. When an LLM generates an incorrect treatment recommendation, no audit trail identifies which training data, which probability weights, or which inference step produced the error - a characteristic that the technical literature terms the “black-box problem.” Existing clinical liability law, built on the premise that the proximate cause of a clinical error can be identified and attributed to a responsible actor, is structurally ill-equipped to handle failures whose causal chain is distributed across a model architecture that no single party fully controls or can fully inspect. Until legal doctrine catches up with this technical reality, the human-in-the-loop requirement functions not only as a patient safety standard but as the primary mechanism by which identifiable legal accountability is preserved: the attending clinician who reviews, authorizes, and documents an AI-informed recommendation remains, under current frameworks, the legally identifiable proximate decision-maker. The FDA’s SaMD framework, published in 2024, classifies adaptive clinical AI including LLMs as the most regulated category of software medical device, requiring pre-market review for high-risk clinical applications and continuous post-market performance monitoring [81]. The EU AI Act of 2024 classifies medical decision-support AI as high-risk and imposes transparency, oversight, and surveillance requirements [38]. ESMO's ELCAP guidance by Verlingue and colleagues articulates professional consensus principles for transparent limitations disclosure and mandatory human oversight in oncological LLM use [83]. Until legal frameworks mature, AI tools should be deployed in roles where the attending clinician's independent judgment is explicitly preserved and documented as the proximate clinical decision, with AI outputs recorded as informational inputs rather than autonomous recommendations.

Table 9. Regulatory and Legal Framework for LLM Deployment in Clinical Oncology

Framework and Regulation	Jurisdiction	LLM Classification	Key Requirements for Clinical LLM Deployment
FDA SaMD Framework [2024] [81]	United States	Adaptive algorithm; high-risk SaMD category	Pre-market review for high-risk clinical applications; post-market performance monitoring; Predetermined Change Control Plan for model updates; 510(k) or De Novo classification pathway
EU AI Act [2024] [38]	European Union	High-risk AI system	Transparency requirements; human oversight mandate; accuracy documentation; post-market surveillance; GDPR alignment; creates de facto human-in-the-loop mandate
GDPR (2016/679) [104]	European Union	Data processor and controller regulations	Explicit legal basis for patient data processing; data subject rights to explanation and contestation; restrictions on extra-EEA data transfers; data processing agreement requirement
HIPAA	United States	Covered entity and business associate	Business Associate Agreements with AI vendors; data minimization in API calls; audit trails for AI-assisted decisions; de-identification standards for LLM inputs
ESMO ELCAP Guidance [83]	International (Oncology Professional)	Professional consensus governance standard	Transparent AI limitations disclosure; mandatory human oversight; systematic outcome evaluation; oncology community consensus for responsible LLM integration
IMDRF Guidance	International	Software as a Medical Device	Analytical validity; clinical validity; overall performance demonstration as a continuous obligation, not a one-time pre-market requirement
Legal Liability Analysis (Shumway and Hartman [82])	United States (analysis)	Currently unresolved	Proposed tiered liability model: responsibility allocated proportionally among developers, institutional deployers, and attending clinicians; attending clinician remains proximate legal decision-maker under current frameworks

Abbreviations: SaMD = software as a medical device; PCCP = Predetermined Change Control Plan; FDA = U.S. Food and Drug Administration; EU AI Act = European Union Artificial Intelligence Act; GDPR = General Data Protection Regulation; HIPAA = Health Insurance Portability and Accountability Act; IMDRF = International Medical Device Regulators Forum; ESMO ELCAP = Ethical and Legal Considerations in AI for Patients; BAA = Business Associate Agreement; DPA = Data Processing Agreement; DPIA = Data Protection Impact Assessment.

Table 9. Regulatory and Legal Framework for LLM Deployment in Clinical Oncology

Framework and Regulation	Jurisdiction	LLM Classification	Key Requirements for Clinical LLM Deployment
FDA SaMD Framework [2024] [81]	United States	Adaptive algorithm; high-risk SaMD category	Pre-market review for high-risk clinical applications; post-market performance monitoring; Predetermined Change Control Plan for model updates; 510(k) or De Novo classification pathway
EU AI Act [2024] [38]	European Union	High-risk AI system	Transparency requirements; human oversight mandate; accuracy documentation; post-market surveillance; GDPR alignment; creates de facto human-in-the-loop mandate
GDPR (2016/679) [104]	European Union	Data processor and controller regulations	Explicit legal basis for patient data processing; data subject rights to explanation and contestation; restrictions on extra-EEA data transfers; data processing agreement requirement
HIPAA	United States	Covered entity and business associate	Business Associate Agreements with AI vendors; data minimization in API calls; audit trails for AI-assisted decisions; de-identification standards for LLM inputs
ESMO ELCAP Guidance [83]	International (Oncology Professional)	Professional consensus governance standard	Transparent AI limitations disclosure; mandatory human oversight; systematic outcome evaluation; oncology community consensus for responsible LLM integration
IMDRF Guidance	International	Software as a Medical Device	Analytical validity; clinical validity; overall performance demonstration as a continuous obligation, not a one-time pre-market requirement
Legal Liability Analysis (Shumway and Hartman [82])	United States (analysis)	Currently unresolved	Proposed tiered liability model: responsibility allocated proportionally among developers, institutional deployers, and attending clinicians; attending clinician remains proximate legal decision-maker under current frameworks

Abbreviations: SaMD = software as a medical device; PCCP = Predetermined Change Control Plan; FDA = U.S. Food and Drug Administration; EU AI Act = European Union Artificial Intelligence Act; GDPR = General Data Protection Regulation; HIPAA = Health Insurance Portability and Accountability Act; IMDRF = International Medical Device Regulators Forum; ESMO ELCAP = Ethical and Legal Considerations in AI for Patients; BAA = Business Associate Agreement; DPA = Data Processing Agreement; DPIA = Data Protection Impact Assessment.

8.4. Informed Consent and the Therapeutic Relationship

Most patients do not know whether AI systems contribute to the analysis of their imaging, the interpretation of their molecular profiling results, or the clinical decision support received by their oncologist. The principle of transparency, articulated in ESMO's ELCAP guidance by Verlingue and colleagues, implies that patients should be informed when AI systems contribute materially to their care decisions [83]. Sittig and Singh, in their recommendations for ensuring safety of AI in clinical care published in JAMA, emphasized the therapeutic relationship as a dimension of AI deployment ethics that resists quantification yet exerts documented effects on treatment adherence, psychological wellbeing, and clinical outcomes [80]. AI deployment that increases rather than decreases the time available for human clinical contact is the implementation vision articulated by ASCO's leadership and reported by Cavallo in The ASCO Post, and the oncology community should hold it as the governing standard for evaluating the ultimate clinical value of LLM integration [37].

8.5. Regulatory Frameworks, Clinical Governance, and the Human Experience

The formal registration of LLM-based clinical tools as Software as a Medical Device (SaMD) is no longer a mere abstract regulatory ambition but an active, high-stakes process. The FDA’s Digital Health Center of Excellence has articulated a rigorous, staged pre-market pathway through which AI/ML-enabled software must advance, ranging from intensive pre-submission consultation to 510(k) clearance or De Novo classification (11, 81). This determination depends heavily on the specific risk profile of the intended clinical use, particularly as regulators refine action plans for high-risk generative models (81, 83).

A cornerstone of this modern regulatory approach is the Predetermined Change Control Plan (PCCP). Since the defining characteristic of LLMs is their capacity to update parameters and refine weights through iterative learning, the PCCP allows these systems to remain operational without triggering a disruptive new pre-market submission for every minor modification (81, 83). This plan defines permissible categories of change and outlines performance monitoring conditions to ensure patient safety remains uncompromised (21, 81). Under the International Medical Device Regulators Forum (IMDRF) guidance, evaluation is a continuous obligation, requiring rigorous demonstration of the software’s analytical validity, clinical validity, and overall performance (21, 38).

For oncology institutions, the adoption of these tools necessitates a strict due diligence process (40, 46). Centers that treat SaMD registration as a simple vendor responsibility rather than a shared governance obligation accept a significant degree of liability. The legal boundary between vendor product defect and institutional medical malpractice remains a critical "grey zone." As AI systems become more autonomous, the traditional "learned intermediary" doctrine is being challenged, leaving both clinicians and institutions vulnerable to complex litigation regarding the ultimate burden of AI-assisted clinical errors [82,83,84].

Transitioning from the regulatory to the operational, the implementation of multimodal LLMs introduces a critical layer of user experience (UX) that addresses physician burnout. The modern oncologist is frequently burdened by the need to synthesize disparate inputs-high-resolution imaging, longitudinal pathology, and complex genomics-a process that contributes significantly to cognitive fatigue (45, 71). From a UX perspective, multimodal AI acts as a sophisticated cognitive filter, utilizing cross-modal attention mechanisms to distill complex streams into a prioritized overview [62,63,64]. This shifts the clinician’s role from manual data aggregation to expert curation, providing a streamlined dashboard for multidisciplinary tumor boards and facilitating precision oncology in a time-sensitive manner (58, 72-74).

Finally, the clinical efficiency gained by the physician is mirrored by the profound psychological benefits for the patient. Cancer patients who encounter complex reports they cannot decode often experience significant anxiety proportional to the uncertainty produced by the unreadable text [47]. Low health literacy is a documented predictor of reduced agency and disengagement from shared decision-making [48]. When LLM-based tools enable patients to correctly interpret status markers or grading designations, the benefit operates at both informational and affective levels (49, 54). A patient who understands their simplified report before a consultation is empowered to engage in active comprehension rather than anxious uncertainty, leading to more productive dialogue and improved recall [50,51,52]. Consequently, evaluation frameworks for LLMs in oncology must move beyond factual accuracy to encompass pre-appointment anxiety indices and patient adherence, ensuring the technology supports the emotional trajectory of the patient as effectively as the clinical workflow (12, 55, 56).

9. Domains Requiring Further Development

9.1. Autonomous Clinical Reasoning Under Conditions of Uncertainty

Kresevic and colleagues, publishing in Nature Medicine, evaluated multiple LLMs using curated real patient cases and found that current LLMs do not accurately diagnose patients across all pathologies, perform significantly worse than physicians in realistic clinical scenarios, fail to adhere systematically to diagnostic or treatment guidelines, and cannot reliably interpret laboratory results in clinical context [106]. Cabral and colleagues' Script Concordance Testing demonstrated substantially lower adaptive reasoning performance than knowledge examination performance [86]. Lee, Bubeck, and Petro, publishing in the New England Journal of Medicine, articulated the benefits, limits, and risks of GPT-4 as a clinical AI chatbot and identified adaptive clinical reasoning as the most difficult performance dimension to optimize [107]. These findings define the performance gap that must be closed before any form of autonomous clinical reasoning can be considered for consequential oncological decisions.

9.2. Intraoperative and Procedural Applications

Real-time intraoperative AI assistance in surgical oncology and interventional radiology requires convergence of computer vision, robotics, and generative language reasoning in architectures that have not yet achieved clinical deployment scale. Hashimoto and colleagues, reviewing AI in surgery in Annals of Surgery, identified real-time intraoperative AI as the highest-value and most technically challenging surgical AI application category, requiring coordination of perceptual, decisional, and physical capabilities that no current system integrates [108]. LLM contributions to procedural oncology are currently limited to preoperative planning documentation, perioperative communication, and postoperative summary generation, all text-based tasks distant from the spatial and haptic reasoning demands of intraoperative guidance.

9.3. Validated Multilingual and Culturally Adapted Deployments

As established in Section 7, deployment of LLMs in non-English clinical environments requires dedicated local validation studies, population-specific knowledge base development incorporating regional variant databases and national reimbursement criteria, and cultural adaptation beyond translation. These prerequisites have not been met for any LLM currently in broad clinical deployment in non-English-language oncology contexts, creating a research gap with direct implications for health equity. Zhu and colleagues’ systematic review confirmed that no current general-purpose LLM has been prospectively validated for clinical use in agglutinative-language or other morphologically complex healthcare environments [95]. Addressing this gap is not a peripheral refinement but a precondition for equitable global access to AI-assisted oncological care.

10. Toward Trustworthy AI in Oncology: Implementation Framework and Future Directions

Trustworthy AI in oncology, as articulated by Sittig and Singh in JAMA, encompasses accuracy, transparency, interpretability, fairness, privacy protection, robustness, and appropriate human oversight [80]. The single most clinically consequential distinction between an AI tool that is trustworthy and one that merely appears trustworthy is whether its outputs are grounded in, and traceable to, authoritative guideline sources - or whether they are generated from unconstrained statistical sampling of a training corpus. A grounded system does not offer an opinion on first-line therapy in a given indication: it retrieves the current NCCN or ASCO recommendation, presents it with a verifiable citation, and explicitly flags when guideline coverage is absent. An ungrounded system generates a fluent, authoritative-sounding answer that may reflect a superseded trial, an incorrect dosing schedule, or a combination that exists nowhere in the approved literature. This distinction is invisible to the clinician who does not know to ask - and that invisibility is precisely the patient safety risk that RAG grounding is designed to eliminate. In oncological practice, trustworthiness is therefore most operationally achieved through RAG grounding in institutional authority: LLMs whose responses are constrained by, traceable to, and verifiable against the outputs of authoritative oncological bodies such as ASCO, NCCN, and ESMO. This distinction is invisible to the clinician who does not know to ask, and that invisibility is precisely the patient safety risk that RAG grounding is designed to eliminate. An institutional implementation framework is consolidated in Table 10.

The ASCO Guidelines Assistant, as reported by Cavallo, represents the most mature current operational expression of this architecture [37]. Bedi and colleagues identified the gap between demonstrated technical performance and demonstrated patient-centered clinical benefit as the most urgent challenge for the field [87]. Longhurst and colleagues, publishing in NEJM AI, proposed a network of AI implementation science centers for evaluating clinical LLM effectiveness, providing a framework for closing this evidence gap through prospective clinical trials [109]. Federated learning architectures, as demonstrated by Dayan and colleagues in Nature Medicine for multi-institutional clinical outcome prediction, offer a pathway toward oncology-specific LLMs trained on real-world clinical experience across multiple institutions without centralization of sensitive patient data, potentially improving performance for underrepresented populations [110]. Tu and colleagues, reporting on the trajectory toward generalist biomedical AI in NEJM AI, identified multimodal integration as the defining frontier of next-generation clinical AI [21]. Rajpurkar, Chen, Banerjee, and Topol, reviewing AI in health and medicine in Nature Medicine, identified the human-AI complementarity model, in which the LLM contributes encyclopedic breadth and data integration at scale while the oncologist provides contextual judgment, ethical reasoning, and relational capacity, as the defining design principle for the next generation of clinical AI systems [111]. The proposed model for Human-AI complementarity, where administrative and data-heavy tasks are delegated to AI while physicians focus on ethical judgment and empathetic care, is summarized in Figure 5.

11. Discussion

The evidence reviewed in this manuscript indicates that the clinical value of large language models in oncology depends strongly on the task for which they are used. Applications such as literature synthesis, documentation support, report simplification, and patient communication are closer to routine implementation, whereas autonomous diagnostic reasoning, treatment selection, and complex multimodal decision-making remain insufficiently validated. Clinical fluency should therefore not be interpreted as equivalent to clinical competence.

A major finding of this review is that trustworthy deployment depends not only on model performance but also on system architecture. Retrieval-augmented generation can improve reliability by grounding outputs in current and verifiable oncology guidelines, while multimodal systems offer the potential to integrate imaging, pathology, genomics, and clinical data. However, these technologies should currently be considered decision-support tools rather than substitutes for multidisciplinary clinical judgment.

Important limitations remain, including hallucination, algorithmic bias, privacy concerns, regulatory uncertainty, and reduced generalizability across languages, populations, and healthcare systems. Most available evidence is also based on retrospective datasets, benchmark questions, or technical performance metrics rather than prospective clinical outcomes. Future studies should therefore focus on real-world effectiveness, patient-centered outcomes, workflow efficiency, and local validation. Overall, the most realistic pathway is progressive integration of LLMs into oncology workflows with transparent limitations, continuous monitoring, and mandatory human oversight.

12. Conclusions

The period inaugurated by November 30, 2022 has established LLMs as a transformative presence in oncology whose clinical implications are too well evidenced to be dismissed and too incompletely realized to be deployed without careful governance. Three broad conclusions are drawn from the evidence reviewed in this manuscript.

In specific and well-defined domains, including literature synthesis, clinical documentation, and the simplification of medical information to improve patient health literacy as demonstrated across multiple systematic reviews, LLMs have demonstrated clinically meaningful performance that can be responsibly integrated into oncological practice with appropriate oversight. The evidence on radiology and pathology report simplification, LLM-assisted documentation and time savings, and evidence extraction from unstructured clinical text supports this conclusion and justifies immediate institutional attention to AI adoption strategies in these domains [49].

In domains including RAG-grounded guideline-based decision support and psychosocial adjunct care, the evidence justifies continued investment in controlled clinical evaluation and cautious institutional deployment, provided systems are designed with RAG grounding in NCCN and ASCO guideline corpora, transparent limitations disclosure, mandatory human-in-the-loop review, and systematic outcome monitoring [80]. The expanded MLLM vision, integrating radiology, WSI analysis, radiomics, NGS, transcriptomics, and RWE into unified precision oncology reasoning, is architecturally feasible and clinically compelling but requires prospective clinical trials before broad deployment can be recommended (62, 72, 73, 87). The challenges of multilingual and culturally adapted deployment - encompassing local reimbursement framework integration, population-specific variant databases, and culturally competent communication norms - require dedicated local validation and population-specific knowledge base development as prerequisites for equitable access to AI-assisted oncological care globally [95,96,97,98,99,100,101].

In domains including autonomous adaptive clinical reasoning, real-time procedural assistance, and complex multimodal integration without human review, the technical, regulatory, and evidentiary foundations have not yet been established. The legal liability framework for AI-assisted clinical errors remains unresolved under current SaMD regulatory frameworks in both the United States and the European Union (38, 81, 82). Algorithmic bias toward underrepresented populations demands active demographic performance monitoring as a condition for ethical deployment (102, 103). Data privacy obligations imposed by HIPAA, GDPR, and analogous national data protection frameworks require formal institutional compliance assessment before any cloud-based LLM deployment in clinical settings [105].

The future of oncology will not be characterized by AI replacing human judgment. It will be characterized by human judgment, augmented by rigorously grounded and transparently governed AI, reaching further, faster, and more equitably than it could reach alone. The oncologist who understands the distinction between static parametric knowledge and RAG-grounded dynamic evidence retrieval, who recognizes that probabilistic generation is a solvable architectural challenge addressed by grounding in NCCN and ASCO guidelines, who appreciates that GDPR, HIPAA, and FDA SaMD frameworks create substantive clinical governance obligations, and who holds in mind that the therapeutic relationship remains the proximate determinant of patient experience and outcomes, is the clinician who will extract the greatest clinical value from this technology and identify its limits before those limits cause harm.