Cancer Clinical Decision Support AI

Cancer treatment usually requires a team of specialists meeting weekly. This AI version gives oncologists a team on demand—six agents, live evidence, and one grounded recommendation per case

Comment

Inspiration

I'm a practicing hematology–oncology physician. The hardest part of my job isn't the cases I see in tumor board (A Meeting where clinical decision support is made), it's the cases I see between tumor boards. A new biopsy on a Tuesday morning, an outside patient transferring in, a regimen choice for someone whose comorbidities don't match the trial population. The full multidisciplinary team isn't in the room for those moments. Most of the time it's just me, the chart, and whatever I can pull up from PubMed in the ten minutes before clinic starts.

A real tumor board pulls together a radiation oncologist, medical oncologist, surgeon, pharmacist, pathologist, and molecular specialist because no single oncologist holds all the relevant evidence in their head. I wanted that team available to me, and to any oncologist, on demand, grounded in the actual literature for the case in front of us, not in what I happened to read last quarter.

What it does

A practicing oncologist pastes a clinical case. Six specialist AI agents — radiation oncologist, medical oncologist, surgical oncologist, clinical pharmacist, molecular oncologist (activates only when biomarkers are present), and pathologist (activates only when the diagnosis is ambiguous) — independently research the case in real time across ten retrieval sources: PubMed, Europe PMC, ClinicalTrials.gov, openFDA, DailyMed, RxNorm, CIViC, Semantic Scholar, and a web fallback.

Every agent operates under a hard rule: every clinical claim must be backed by a citation from a paper retrieved during this session. No training-knowledge answers, no (judgment) annotations, no "in my experience" filler. If an agent can't find evidence, it abstains rather than guess.

The agents then debate across rounds. A board-chair LLM judges consensus after each round; the loop continues until they converge or hit a safety cap. The treating oncologist watches the deliberation happen live, then reads a bottom-line synthesized recommendation — citations linked to the source paper, evidence panel filterable by paper type, every claim traceable.

The goal isn't to replace the tumor board or the oncologist. It's to give the oncologist the team and the evidence on demand, in the ten minutes before clinic starts.

Built With

Share this project:

Updates