AI Tumor Board

What this is

AI Tumor Board is an open-source demo of a multidisciplinary oncology board run by up to nine specialist AI agents — a radiation oncologist, medical oncologist, surgical oncologist, clinical pharmacist, molecular oncologist, pathologist, hematologist, transplant / cellular-therapy specialist, and a clinical-trial matcher. Paste a clinical vignette and watch the agents independently research the case across PubMed, ClinicalTrials.gov, openFDA, DailyMed, RxNorm, CIViC, and the open web, then discuss across rounds until they converge on a consensus recommendation — which you can export to PDF.

Every recommendation is grounded in evidence retrieved at run-time. Agents that cannot find supporting evidence abstain rather than answer from training knowledge. Citations link to the underlying source.

How it works

• Nine specialist agents, each on a frontier LLM with a specialty-specific tool subset and PubMed MeSH bias.
• Five are conditional — they decide for themselves whether the case is in their lane and self-skip when it isn't: the molecular oncologist (needs biomarker data), the pathologist (diagnostic ambiguity), the hematologist (a hematologic malignancy or heme emergency), the transplant / cellular-therapy specialist (when stem-cell transplant or CAR-T is in play), and the clinical-trial matcher.
• Shared findings: the molecular, pathology, and hematology agents broadcast their findings to the rest of the board, so downstream specialists (e.g. transplant) reason from the established diagnosis and risk category.
• Strict evidence-only rule: every clinical claim must be backed by a [N] citation. (judgment) annotations are not accepted.
• Discussion loop: round 1 is independent research; subsequent rounds prepend the other specialists' positions plus the board chair's open questions. The loop continues until consensus (alignment ≥ 0.85) or the round cap (2 by default).
• Live streaming UI via Server-Sent Events — you watch each agent's tool calls, drafts, and the chair's verdict as they happen.
• Export: download the final consensus recommendation — with the case, inline citations, and evidence list — as a PDF.

Tech stack

• Backend: FastAPI, a frontier LLM via an OpenAI-compatible client (OpenAI or Google Gemini, provider-configurable), Biopython, httpx
• Frontend: vanilla HTML / CSS / JS (no framework), marked.js for markdown
• Streaming: SSE via FastAPI StreamingResponse
• Sources: NCBI Entrez (PubMed), Europe PMC, Semantic Scholar, ClinicalTrials.gov v2, openFDA, DailyMed, RxNorm, CIViC GraphQL, Brave Search

Source code: github.com/Roupen92/AI-tumor-board

About the author

Dr. Roupen Odabashian

Dr. Roupen Odabashian is a hematology–oncology physician and the founder of MeDucation AI, a platform reimagining how the next generation of oncologists learn. His clinical work at the bedside — where treatment decisions hinge on rapidly evolving evidence, complex genomics, and the realities of individual patients — is the foundation for everything he builds. He created this AI Tumor Board as an educational space where trainees, fellows, and practicing clinicians can work through real oncology cases alongside AI, sharpening clinical reasoning outside the time pressure of a live conference room.

A practicing physician and an active voice in healthcare AI, Roupen works at the intersection of medicine and technology — translating what bedside oncology actually needs into tools clinicians will actually use. He hosts the Delta HealthTech podcast, where he sits down with the founders, researchers, and clinicians shaping the future of healthcare, and he writes and speaks regularly about how large language models are reshaping clinical education and decision-making.

His mission is simple: make oncology learning faster, deeper, and more accessible, and put modern AI in the hands of every clinician who treats cancer.

GitHub · Email