The AI-based Tumor Board: A Multi-agent approach to finding cancer treatments with AI.

Every week, 30 of Europe’s leading experts in Leukemia and Lymphoma gather online for the International Leukemia/Lymphoma Tumor Board (iLTB). Their mission is crucial: to tackle some of the most complex and treatment-resistant pediatric leukemia and lymphoma cases across Europe. This multidisciplinary team is composed of specialists in immunotherapy, genetics, molecular biology, clinical trials, and cellular therapies, from all around Europe, who work together to identify life-saving treatment plans for these children.

During these meetings, physicians present their patient’s profiles, including details on prior treatments, relapses, genetic findings, and immunophenotyping results. Given that many patients already have undergone multiple lines of therapy, figuring out the optimal treatment plan requires the combined knowledge of the panel of experts. After each case presentation, the board engages in an in-depth discussion to come up with a personalized and evidence-based treatment plan for each patient. These decisions aren’t just academic exercises — they’re often the last hope for children who have run out of treatment options.

To support the iLTB, ML6 developed the AI-based Tumor Board (AITB). The AITB is an AI-powered tool that uses specialized AI agents to analyze and discuss patient’s cases prior to the actual Tumor board meeting. By providing each AI agent with pre-screened scientific literature and clinical trial data for each case, the AITB can simulate case discussions and produce a detailed report highlighting key discussion points, insights, and potential treatment options. By offering an overview of available therapies and rationale before the live case review, the AITB aims to improve the efficiency and effectiveness of the tumor board’s decision-making process, potentially leading to better decision-making and possibly better patient outcomes.

The integration of the AI-based Tumor Board into the iLTB’s workflow could offer significant advantages. By leveraging AI to go through academic papers and the latest clinical trial data, the AITB ensures that treatment recommendations are grounded in the most recent evidence. Additionally, the AITB can introduce new perspectives by suggesting treatment options that might otherwise be overlooked or not immediately apparent to the specialists. Finally, the AI-led TB allows the iLTB expert panel to better prepare for each meeting and save precious time, allowing them to focus on what truly matters: making life-saving decisions for their patients

How Does the AI-based Tumor Board Work?

‍

At its core, the AI-led Tumor Board operates as a multi-agent system — think of it as a team of virtual specialists, mirroring the roles of human experts in the actual tumor board. Using knowledge from medical literature, the AI agents work together to analyze a patient’s case. But before we explore how these AI agents collaborate, let’s start by understanding what defines a single AI agent.

Let’s Start with a Single AI agent

‍

To understand the AI-led Tumor Board (AITB), we first need to explore how a single AI agent operates. At its core, an AI agent is powered by a large language model (LLM) like ChatGPT, Gemini, or others — but it’s much more than just an LLM. Unlike traditional LLMs, AI agents are autonomous entities that perceive their environment and can take actions to achieve specific goals.
In the context of the AITB, each agent is equipped with three key capabilities: Knowledge, memory, and tools.

• Knowledge: Unlike the static, general knowledge embedded in traditional LLMs, each agent in the AITB has access to a specialized, pre-screened literature database. For instance, the genetics agent has its dedicated selection of papers related to only genetic alterations such as KRAS or KMT2A fusions. This specialization ensures that the agents’ responses are deeply grounded in relevant scientific literature, significantly reducing the risk of hallucinations.
• Memory: By enabling memory, AI agents can retain context of discussions, keeping track of what has already been discussed. This feature is essential for collaborative, multidisciplinary thinking, ensuring that insights from one part of the conversation inform later stages.
• Tools: AI agents can iteratively search through their knowledge bases based on what was said. In the future, this toolset could expand to include real-time access to external sources like PubMed and Google Scholar.

This approach addresses the limitations of standard LLMs, enabling AI agents to provide specialized and well-grounded medical knowledge that isn’t typically present in standard language models like ChatGPT. For a more in-depth explanation about agents, refer to our blogpost: Unlocking the Power of AI Agents.

But this only describes a single AI agent, what about the multi-agent system?

Why a multi-agent approach?

‍

Instead of using a single AI agent, the system uses a multi-agent conversation between distinct specialized AI agents.

Some of you might be wondering:

Why do we need a multi-agent conversational framework with multiple AI agents? Why not just feed all the papers into a single, “all-knowing” agent and ask it to list all treatment options and clinical trials?

Fostering collaborative discussions

Much like the traditional tumor board — where specialists from different fields collaborate to decide on a patient’s treatment — a multi-agent system brings multiple perspectives to the table. This ensures that critical factors like drug interactions, patient history and emerging clinical trials are not overlooked. Relying solely on a single LLM can limit the depth and nuance of these discussions, resulting in one-size-fits-all responses.

Improving accuracy and reliability

Additionally, multi-agent systems can significantly enhance the accuracy and reliability of recommendations. While general-purpose LLMs are powerful, studies show that they can produce hallucinations, biased content or irrelevant information (Haltaufderheide & Ranisch, 2024; Du et al., 2023; Ahmad et al., 2023; Ray et al., 2024; Li et al., 2024). By distributing reasoning tasks among specialized agents — each focused on different facets of the problem — errors are minimized. Agents can cross-check each other’s insights, refining recommendations to ensure they are accurate and valid. Additionally, they can revisit earlier statements, update them with new evidence, and resolve contradictions — an essential part of fruitful discussions.

In our case, agents can compare the efficacy of a given treatment or treatment combination by comparing the overall survival data from a given paper. In addition, the AITB can cross-check for treatment side effects and restrictions based on the patient history.

The AI-led Tumor Board in Action

‍

Let’s see how such a conversation can look like.

In our demo, the cellular therapy expert begins by querying its knowledge base and suggests possible treatments tailored to the patient’s unique medical profile. Importantly, each proposal is grounded is by referencing a specific page from a paper in the knowledge base. This allows the human experts to validate the information if needed. Also note that we specifically instructed the agents to focus on survival rates and event-free rates (indicating that the disease has not returned), as these are critical factors in determining the most effective treatment option for the patient.

Next, the genetic expert follows up with additional insights about the KRAS mutation and additional biomarkers.

‍

The conversation continues.

The agents make use of toolcalls that allow them to ground their responses in a selection of academic papers.

Once the debate concludes, the AITB generates a detailed report that summarizes the experts’ insights and outlines the recommended treatment plans. Additionally, the system matches the proposed treatments with relevant clinical trials based on the patient’s biomarkers.

‍

Ethical considerations, legislations and risks

‍

This blog post serves as a demo and Proof of Concept (PoC) to showcase the potential of the AITB. Note that it contains unverified medical data based on a fictitious scenario and has not undergone rigorous accuracy checks. As such, any conclusions presented here should be interpreted with caution and must not be used for clinical decision making or patient care.

Developing AI tools for healthcare requires careful consideration of risks such as false or biased outputs and a critical examination of when and how AI can assist medical experts responsibly. However, risks can be mitigated by grounding the system’s output in verified medical literature, designing systems that allow easy fact-checking, and requiring validation by experts. Additionally, proper safeguards and user education must be implemented to ensure responsible usage and prevent misuse.

At ML6, we have several initiatives advancing responsible AI. Interested? Check out our blog posts by the Ethics Special Force: The Journey Towards Responsible AI at ML6; Implementing AI Governance: A Focus on Risk Management; Navigating High-Risk AI Systems under the European AI Act: a Guide for Early Stages.

The future of the AI-based Tumor Board

‍

By integrating the AI-led Tumor Board into the iLTB’s workflow, the medical experts can come better prepared and save valuable time before the live case discussions.

Future iterations could incorporate real-time web searches, automated fact-checking and even personalized AI agents. Imagine each expert having a digital twin that stays up-to-date with what they’ve recently read and studied. Imagine AI agents continuously updating themselves by pulling in new research as it becomes available, while also learning from expert feedback and the patient’s response and side effects to the suggested treatment. What if, instead of separate pre-meeting discussions, AI agents could actively participate in the iLTB sessions, stepping in when a critical insight is being overlooked? Sounds exciting, right?

The potential of AI-powered multi-agent systems like the AI-led Tumor Board extends far beyond healthcare. Any field that relies on expert collaboration — such as life sciences, education or finance — could be transformed by this technology.

While we must remain mindful of ethical considerations and the importance of data grounding, the rapid advancement of AI technology presents unprecedented opportunities. And while some ideas may seem futuristic, the time to experiment is now.

‍

This blogpost was written in collaboration with Uri Ilan. MD, lead of innovation at the Princess Maxima Centrum

‍

Watch the full demo here.