Can Artificial Intelligence Help Cancer Tumor Boards? A Comprehensive Review of Large Language Models in Cancer Care

Table of Contents

• Key Points
• Background: Why This Research Matters
• How the Research Was Conducted
• What Studies Were Included
• Key Findings: How Well Did AI Perform Across Cancer Types?
• Performance by Type of Clinical Task
• The Impact of Model Architecture and Evolution
• Clinical Implications: What This Means for Patients
• Limitations: What the Study Could Not Prove
• Recommendations and Next Steps
• Frequently Asked Questions
• Source Information

Key Points

• A review of 31 studies covering 3,845 patient cases found large language models are not yet reliable enough for independent cancer treatment decisions without human oversight.
• In common guideline-driven cancers such as breast and prostate, models reached up to 93.9% and 93% agreement with human tumor boards respectively.
• Performance dropped in rare or complex cancers such as sarcoma and neuro-oncology, and models cannot interpret medical images, relying on text reports.
• Models can hallucinate false information and showed overtreatment bias in ambiguous neuro-oncology cases, so clinician verification is required.
• The authors recommend human-in-the-loop safeguards, multimodal architectures, and rigorous prospective validation before clinical adoption.

Background: Why This Research Matters

Multidisciplinary tumor boards (MDTs) are the gold standard for cancer care. At these meetings, specialists from surgery, medical oncology, radiation oncology, pathology, and radiology come together. They review each patient's case and formulate evidence-based treatment recommendations.

These collaborative forums face increasing pressure. Cancer incidence is rising. Treatment complexity is growing. Molecular data requiring interpretation is expanding rapidly. The integration of precision medicine has further complicated decision-making. Molecular tumor boards now must interpret genomic alterations, identify targeted therapies, and match patients to clinical trials.

To address this increasing complexity, researchers initially turned to classical Natural Language Processing (NLP) — computer systems that can read and structure text. These tools helped optimize tumor board workflows through data structuring and aggregation. Before generative LLMs became widespread, tools like the "AI Pathway Companion" showed the potential of NLP-driven Clinical Decision Support Systems (CDSS) to enhance efficiency. Studies demonstrated that these systems could significantly reduce case preparation times and improve clinician satisfaction in prostate cancer pathways. They did this by automatically extracting and presenting fragmented data from disparate clinical systems.

Building on these foundations, Large Language Models (LLMs) have emerged as the next generation of tools. LLMs offer capabilities in natural language processing, pattern recognition, and knowledge synthesis that could augment human decision-making. Recent advances in transformer-based architectures — exemplified by GPT-4, Claude-3, and open-source alternatives like Llama — have demonstrated remarkable abilities. These models can process medical literature, interpret clinical guidelines, and generate coherent clinical recommendations.

However, applying these general-purpose models to the highly specialized domain of oncology tumor boards raises critical questions. These questions concern accuracy, safety, and clinical utility.

The potential benefits of LLM integration in tumor boards are substantial. These systems could:

• Standardize decision-making across institutions
• Reduce preparation time for complex cases
• Ensure comprehensive consideration of treatment options
• Improve adherence to evolving guidelines

For resource-limited settings, locally deployed open-source models might democratize access to expert-level decision support. Additionally, LLMs could facilitate clinical trial matching. This is a labor-intensive process that currently results in only 3%–5% enrollment rates despite many eligible patients.

However, significant concerns temper this enthusiasm. The "black box" nature of LLM decision-making — meaning it is often unclear how the model reaches its conclusions — poses a barrier. The potential for hallucinations (the model generating false or fabricated information) is another concern. Reliance on potentially biased training data and the medicolegal implications of AI-assisted recommendations all pose barriers to implementation. Furthermore, the dynamic nature of oncology challenges models trained on historical data. Treatment paradigms evolve rapidly, and practice-changing trials are published continuously.

This systematic review addresses the current applications of LLMs in tumor board decision-making. It compares AI-generated recommendations to human MDT decisions. It assesses technical implementation requirements. It also evaluates identified biases and safety concerns to determine the current technology readiness level for clinical deployment.

How the Research Was Conducted

This systematic review was conducted following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines. The review protocol was developed in advance, specifying eligibility criteria, search strategy, and data extraction procedures.

The review was not prospectively registered with PROSPERO or any other systematic review registry. The authors acknowledge this as a limitation. However, all eligibility criteria, search strategies, and data extraction procedures were specified prior to the literature search and followed throughout.

Inclusion criteria encompassed primary research articles published in peer-reviewed journals or high-quality preprints between January 2023 and November 2025. Eligible studies evaluated LLM applications in oncology tumor boards or multidisciplinary team meetings. They also had to report quantitative performance metrics comparing LLM recommendations to human decisions or guidelines. Only English-language publications were considered.

Exclusion criteria removed review articles, editorials, or opinion pieces without original data. Studies using traditional machine learning or rule-based systems without LLM components were excluded. Studies not specifically addressing tumor board or MDT decision-making were excluded. Conference abstracts without full-text availability were also excluded.

The search was conducted across PubMed/MEDLINE, Embase, IEEE Xplore, and arXiv. The search strategy combined three concept groups using Boolean AND operators:

1. Terms related to large language models ("large language model" OR "LLM" OR "ChatGPT" OR "GPT-4" OR "GPT-3.5" OR "GPT-4o" OR "Claude" OR "Gemini" OR "generative artificial intelligence" OR "generative AI")
2. Terms related to tumor boards and multidisciplinary oncology ("tumor board" OR "tumour board" OR "multidisciplinary team" OR "multidisciplinary" OR "MDT" OR "cancer conference" OR "treatment decision" OR "treatment planning")
3. Terms related to oncology ("cancer" OR "oncology" OR "neoplasm" OR "malignancy" OR "tumor" OR "tumour")

The search was restricted to publications from January 2023 to November 2025. No language filter was applied at the search level. Non-English publications were excluded during screening.

Two reviewers independently screened titles and abstracts, followed by full-text review of eligible articles. Disagreements at each stage were resolved through consensus discussion. Data extraction captured study characteristics, cancer types, LLM models, clinical applications, and performance metrics.

Due to substantial heterogeneity in study designs and outcome measures, a narrative synthesis was performed. Studies were grouped by primary application and cancer type.

A quality assessment of included studies was conducted using a purpose-developed framework for LLM evaluation studies. Given the absence of established quality assessment tools for AI-assisted clinical decision-making research, the authors developed a novel ten-domain framework. This framework was informed by principles of diagnostic accuracy assessment (QUADAS-2) but specifically tailored to the methodological challenges unique to LLM evaluation. The framework assesses:

• Case selection method
• Case complexity spectrum
• Prompt reproducibility
• Repeat query testing
• Model parameter reporting
• Assessor blinding
• Independence of the reference standard
• Use of validated outcome metrics
• Error severity classification
• Data realism

As this framework has not been externally validated, the quality assessment results should be interpreted as exploratory rather than definitive. Each domain was rated by the first reviewer and verified by the second reviewer.

What Studies Were Included

The initial database search yielded a total of 741 records. After removing 157 duplicates, 584 records remained for title and abstract screening. Two reviewers independently screened these records and excluded 507 that did not meet the scope of the review. Examples of excluded studies included those focused on traditional machine learning or rule-based systems rather than generative LLMs. Examples also included studies addressing multidisciplinary teams outside of oncological decision-making. Examples also included studies not evaluating LLM performance in a tumor board or MDT context.

The remaining 77 full-text articles were retrieved and assessed for eligibility. Of these, 46 were excluded for reasons including lack of original data (review articles or opinion pieces). Reasons also included absence of an LLM component (traditional machine learning only). Reasons also included lack of specificity to tumor board or MDT decision-making, or being conference abstracts only. Ultimately, 31 studies met the inclusion criteria and were included in this review.

This systematic review analyzed 31 studies published between 2023 and 2025, encompassing a total of 3,845 unique patient cases. The research landscape is global, with a notable concentration of studies in Germany (10 studies), followed by Turkey (4), Italy (4), and China (3).

Study designs were predominantly retrospective comparative analyses. The majority (29 of 31) evaluated LLM-generated recommendations against actual multidisciplinary tumor board decisions. Of note, three included studies were preprints at the time of inclusion and had not undergone formal peer review. These are clearly labeled in the supplementary tables, and their findings should be interpreted with this caveat in mind.

Beyond binary concordance (simple agreement or disagreement), studies employed a range of evaluation instruments. These included:

• The AI Performance Instrument (AIPI)
• Likert-scale assessments for medical adequacy
• Levels of Evidence classification
• Information density metrics
• Consistency analyses across repeated queries

Two studies employed more advanced methods. Schmidl et al. applied Kaplan–Meier and Cox proportional hazards models for survival analysis of LLM-selected versus tumor board–selected treatment cohorts. Li et al. conducted prospective validation through a ring trial across 21 sarcoma centers.

A total of 52 LLM instances were identified across the included studies. An LLM instance is defined as a unique model–study pairing, meaning each distinct model version evaluated within a given study counts as one instance. The OpenAI GPT family was the most extensively assessed (29 instances), spanning from ChatGPT-3.5 to GPT-4o and the reasoning model ChatGPT-o1. The Claude family accounted for 5 instances (Claude-2, Claude 3 Opus, Claude 3.5 Sonnet). The Llama family accounted for 5 instances (Llama 2 through Llama 3.2-vision). Google Gemini accounted for 2 instances.

Notably, specialized systems were also evaluated. These included a guideline-augmented chatbot (TheSerenityBot, Claude-2 based), a trial-matching model (Trial-LLAMA), and the reasoning model DeepSeek-R1 in multi-agent configurations.

Breast cancer was the most frequently studied domain (11 studies), followed by head and neck cancer (6 studies), pan-cancer evaluations (4 studies), sarcoma (2), and molecular tumor board applications (2). The remaining studies addressed one domain each: genitourinary cancer, gynecologic oncology, lung cancer, neuro-oncology, prostate cancer, and general MDT settings.

Key Findings: How Well Did AI Perform Across Cancer Types?

An apparent pattern emerged across clinical domains, which the authors term the "complexity gradient." LLM performance appeared to correlate inversely with the complexity of the clinical decision-making required. In other words, the more complex the decision, the worse the AI performed. However, there was substantial methodological heterogeneity across studies. This heterogeneity included differences in model versions, prompting strategies, and evaluation metrics. This pattern should be interpreted as an observational trend requiring prospective validation rather than an established causal relationship.

In standardized, high-incidence domains with robust guideline-based algorithms, advanced models demonstrated high concordance with human MDTs. In breast cancer, GPT-4o achieved up to 93.9% concordance when using structured prompts and institutional protocols. A prompt-driven AI assistant showed very good agreement for neoadjuvant therapy suitability (κ = 0.91). In prostate cancer, Kaiser et al. reported 93% alignment in pre-therapeutic decision-making using structured clinical data. In radiation oncology queries, ChatGPT-3.5 responses were on par with or superior to expert answers in 94% of cases.

These findings suggest that for cancers with well-defined, algorithm-based guidelines, LLMs can effectively retrieve and apply standard-of-care protocols.

However, this reliability diminished in complex, rare, or nuanced clinical scenarios. In neuro-oncology, Tini et al. found that while ChatGPT-4 aligned well with experts for clear-cut glioblastoma cases, it exhibited systematic overtreatment bias in intermediate "gray zone" scenarios. The model adopted a "when in doubt, treat" approach.

In sarcoma — a heterogeneous group of rare tumors — ChatGPT-4o showed variable performance across specialties with a mean Likert score of 3.76/5. Even advanced models like Claude 3.5 Sonnet and OpenAI-o1 frequently failed to align with expert consensus in a multi-center ring trial. In laryngeal cancer, ChatGPT-4 provided recommendations consistent with MDT decisions in only 72% of patients. In thoracic MDT boards, overall concordance was 76% (κ = 0.59). Across 102 patients with various cancer types, concordance between ChatGPT-4o and tumor board decisions remained low.

This observed "complexity gap" suggests that current LLMs may excel at pattern matching within established guidelines. However, they appear to struggle with the abductive reasoning required for the exceptions and edge cases that typically necessitate MDT discussion.

Beyond disease site, the specific nature of the clinical task influenced model reliability. Treatment recommendation generation was effective in guideline-driven scenarios but degraded for complex or rare presentations. Guideline adherence checking achieved up to 100% accuracy when guidelines were explicitly integrated via retrieval-augmented generation (RAG). Clinical trial matching showed moderate improvement (+5% for molecular alteration extraction). Staging classification depended heavily on the quality and structure of input data. Documentation and summarization tasks leveraged the natural language processing strengths of LLMs most directly.

Several studies identified specific sources of discordance beyond clinical complexity. Wang et al. found that ChatGPT lagged significantly behind physicians in feasibility and appropriateness for ICU oncology patients. This suggests that models struggled with holistic patient assessment beyond guideline application. Umihanic et al. observed that discordance frequently arose from the model's inability to account for patient-specific contextual factors not captured in the structured prompt, such as frailty or local resource limitations.

Additionally, the text-only nature of current LLMs created limitations in inherently multimodal disciplines. Sorin et al. and Li et al. noted that models relying on text reports rather than raw imaging data lacked the granular detail required for surgical planning and margin assessment.

Performance by Type of Clinical Task

The review also examined how LLMs performed across different types of clinical tasks. The findings are summarized below.

Treatment recommendation: LLMs worked well in standard, guideline-driven scenarios (such as breast and prostate cancer) and for high-level decisions (such as curative versus palliative). They struggled with complex or nuanced patient factors, specific choices for advanced or rare cancers (such as sarcoma or second-line head and neck cancer), and invasive procedures. The reasons for these struggles include a lack of deep clinical reasoning and contextual understanding, as well as an inability to interpret multimodal data.

Clinical trial eligibility: LLMs worked moderately well for refining molecular alteration extraction to improve trial matching (+5% improvement). They are good at specific text processing, but overall matching remains complex.

Staging classification: LLMs worked well when provided with structured inputs (such as TNM stages). Performance depends on data quality and structure.

Guideline adherence checking: LLMs worked well when guidelines were explicitly integrated via RAG or structured prompts (up to 100% accuracy). They struggled with reliably citing sources without integration and could exhibit undertreatment bias. RAG and prompt engineering significantly boost retrieval accuracy compared to base model knowledge.

Molecular data interpretation: LLMs worked moderately well for extracting information from text reports. They struggled with complex genomic interactions or rare mutations. High information density handling is good, but biological reasoning is limited.

Documentation and summarization: LLMs worked well for summarizing complex clinical situations and generating sensible decisions. Strong natural language processing capabilities are utilized for administrative efficiency.

Open-source and locally deployable models showed promise in specific niches. Llama-3 achieved 92% concordance for curative versus palliative treatment distinction in head and neck cancer, outperforming ChatGPT-4o in this specific task. This suggests that locally run models may offer viable decision support in certain contexts.

The Impact of Model Architecture and Evolution

The review period (2023–2025) captured a rapid evolution in model capabilities. Earlier studies utilizing GPT-3.5 consistently reported lower concordance and higher hallucination rates compared to later investigations employing GPT-4 and GPT-4o.

Griewing and Knitza documented this trajectory, showing that the upgrade to GPT-4 significantly improved alignment with breast cancer tumor board decisions. Claude-3-Opus demonstrated superior comprehensiveness compared to GPT-4, achieving an F1 score of 0.95 for overall treatment in breast cancer.

By contrast, Google's Gemini showed significantly lower quality assessment scores compared to ChatGPT-4 in head and neck cancer (median QAMAI score 2 vs. 3, p = 0.004). Gemini also showed only slight agreement with MDT decisions in a pan-cancer evaluation.

Response stability also varied across model architectures. Claude-3.5-Sonnet achieved 100% consistency across repeated queries in head and neck cancer cases. This suggests that response stability varies substantially across model architectures and may represent a viable decision-support option in some settings.

Clinical Implications: What This Means for Patients

This review has several important implications for patients and their families.

First, current LLMs are not ready to make cancer treatment decisions on their own. The authors state clearly that these models "are not yet reliable enough for autonomous decision-making." Patients should be reassured that human expertise remains essential in cancer care.

Second, LLMs show real promise as assistive tools. They can help with documentation, summarization, and decision support in routine oncological cases. This means that in the future, AI might help doctors prepare for tumor boards more efficiently, potentially freeing up more time for direct patient care.

Third, the accuracy of AI recommendations depends heavily on the type of cancer. For common cancers with well-established treatment guidelines — such as breast and prostate cancer — AI models can achieve high agreement with human experts (up to 94%). For rare or complex cancers — such as sarcoma or brain tumors — AI performance drops significantly.

Fourth, AI models currently cannot interpret medical images. They rely on text reports rather than raw imaging data. This limits their usefulness in specialties that depend heavily on imaging, such as radiology and surgical planning.

Fifth, AI models may exhibit biases. In neuro-oncology, for example, ChatGPT-4 showed a systematic overtreatment bias in ambiguous cases. This means the model tended to recommend treatment even when the evidence was unclear. Such biases could lead to inappropriate treatment recommendations if not caught by human reviewers.

Sixth, AI models can "hallucinate" — that is, generate false or fabricated information. This is a significant safety concern. It means that any AI-generated recommendation must be carefully verified by a human clinician.

Limitations: What the Study Could Not Prove

The authors acknowledge several important limitations of this review.

The review was not prospectively registered with PROSPERO or any other systematic review registry. While all procedures were specified in advance, this is a limitation.

There was substantial heterogeneity in study designs and outcome measures. This made it impossible to perform a meta-analysis (a statistical combination of results). Instead, the authors performed a narrative synthesis.

The quality assessment framework used in this review has not been externally validated. The results of the quality assessment should therefore be interpreted as exploratory rather than definitive.

Three included studies were preprints at the time of inclusion and had not undergone formal peer review. Their findings should be interpreted with caution.

The observed "complexity gradient" — the pattern where AI performs worse on more complex cases — should be interpreted as an observational trend. It requires prospective validation rather than being considered an established causal relationship.

Finally, the rapid evolution of LLM technology means that findings may become outdated quickly. Models evaluated in 2023 may already be superseded by newer versions.

Recommendations and Next Steps

Based on their findings, the authors make several recommendations for the future of LLMs in tumor boards.

1. Human-in-the-loop safeguards are essential. Any clinical implementation of LLMs must include human oversight. AI should assist, not replace, human decision-making.
2. Multimodal architectures need to be developed. Future models must be able to interpret non-textual data, such as medical images, to be useful in specialties like radiology and surgical planning.
3. Rigorous prospective validation is required. Before LLMs can be widely adopted in clinical practice, they must be tested in prospective studies that demonstrate safety and efficacy.
4. Standardized evaluation frameworks are needed. The authors developed a novel ten-domain framework for this review, but it has not been externally validated. The field needs standardized tools for assessing AI performance in clinical decision-making.
5. Model selection should match the clinical task. Different models perform differently across tasks and cancer types. Institutions should carefully evaluate which models are best suited for their specific needs.
6. Guideline integration improves accuracy. When guidelines are explicitly integrated via retrieval-augmented generation (RAG) or structured prompts, accuracy can reach up to 100%. This suggests that how models are used matters as much as which models are used.

The authors conclude that current LLMs exhibit sufficient maturity to function as assistive tools for documentation and decision support in routine oncological cases. However, they are not yet reliable enough for autonomous decision-making. Successful clinical implementation will require human-in-the-loop safeguards, the development of multimodal architectures, and rigorous prospective validation to ensure patient safety.

Frequently Asked Questions

Can AI make my cancer treatment decisions?

No. In a review of 31 studies covering 3,845 patient cases, the authors concluded that current large language models are not yet reliable enough for independent decision-making without human oversight. Human expertise remains essential. Any clinical use must include human-in-the-loop safeguards, meaning AI assists doctors rather than replacing their judgment about your care.

How accurate was AI compared with human tumor boards?

Accuracy depended on cancer type. In common cancers with clear guidelines, such as breast cancer, GPT-4o reached up to 93.9% agreement with human tumor board decisions using structured prompts. In prostate cancer, one study reported 93% alignment. However, agreement dropped in rare or complex cancers, including sarcoma and neuro-oncology, where models often failed to match expert consensus.

Why does AI perform worse on rare or complex cancers?

The review describes a complexity gradient: the more complex the decision, the worse the AI performed. Models did well matching patterns within established guidelines but struggled with the reasoning needed for exceptions and edge cases. They also cannot interpret medical images, relying on text reports, which limits usefulness in sarcoma, brain tumors, and surgical planning.

What risks or biases did the review find in AI recommendations?

Two safety concerns were identified. Models can hallucinate, generating false or fabricated information, so any AI output must be verified by a clinician. Models can also show bias: in neuro-oncology, ChatGPT-4 showed systematic overtreatment bias in ambiguous gray-zone cases, tending to recommend treatment even when evidence was unclear. Discordance also arose when models missed patient factors like frailty.

What does 94% agreement between AI and experts actually mean?

It means that in certain studies, AI recommendations matched human tumor board decisions in 94% of cases. This was seen in radiation oncology queries, where ChatGPT-3.5 responses were on par with or superior to expert answers. It does not mean the AI was correct in every case, and it does not mean AI can decide treatment alone.

Could AI help match me to a clinical trial?

Possibly, but only as an assistive tool. The review found moderate improvement of about 5% for molecular alteration extraction to support trial matching. Overall matching remains complex. Clinical trial enrollment currently runs only 3% to 5% despite many eligible patients, and AI is not yet reliable enough to handle this independently.

What would AI realistically be used for in cancer care now?

The authors concluded current models are mature enough to assist with documentation and routine decision support, not autonomous decisions. Documentation and summarization tasks suit them well. Guideline adherence checking reached up to 100% accuracy when guidelines were explicitly integrated. Any clinical implementation still requires human oversight, multimodal development, and rigorous prospective validation before wider adoption.

If AI agrees with my tumor board's treatment plan, do I still need a second opinion?

AI models reach high agreement with human tumor boards in common, guideline-driven cancers such as breast and prostate cancer, up to 94%, but agreement drops in rare or complex cancers such as sarcoma and neuro-oncology. Models cannot interpret medical images, may hallucinate false information, and in ambiguous brain tumor cases showed a systematic overtreatment bias. AI is not reliable enough for independent decision-making, so human oversight remains essential. A second opinion from independent experts can review whether a recommendation fits your specific case. Diagnostic Detectives Network provides independent expert second opinions.

Source Information

Original article title: Applications of large language models in tumor boards: a systematic review.

Authors: Konzmann B, Henkel M, Breit C, Wetterauer C.

Affiliations: Institute of Urology, University Hospital Basel, Basel, Switzerland; Digitalisation and ICT, University Hospital Basel, Basel, Switzerland; Institute of Radiology, University Hospital Basel, Basel, Switzerland

Journal: Oncology Reviews, Volume 20, Article 1757059

DOI: 10.3389/or.2026.1757059

Publication date: 20 May 2026

Article type: Systematic Review

Note: This patient-friendly article is based on peer-reviewed research. The original article was published under a Creative Commons Attribution License (CC BY), which permits use, distribution, and reproduction in other forums provided the original authors and copyright owner are credited.