How Can Oncologists Start Experimenting With AI Tools Today?

This segment of the AI Oncology Revolution opened with Arturo Loaiza-Bonilla, MD, MSEd, FACP, breaking down retrieval-augmented generation (RAG), and Matthew Matasar, MD, of Rutgers Cancer Institute, cautioning that not every RAG-based tool on the market is HIPAA compliant. From there, the conversation turned practical: Matasar’s advice for a clinician curious about these tools is simply to take the leap — running a real case through a system like Doc GPT alongside a live tumor board discussion, without ceding any actual decision-making to the model. Matasar floated a prospective trial built on exactly that comparison, pitting tumor boards against generative models and RAG systems and having blinded experts assess the results.

Loaiza-Bonilla is systemwide chief of hematology and oncology at St. Luke’s University Health Network and co-founder of Massive Bio. Matasar is chief of the Division of Blood Disorders at Rutgers Cancer Institute/Jack & Sheryl Morris Cancer Center and professor of medicine at Rutgers Robert Wood Johnson Medical School, where his clinical and research interests focus on lymphoma, and editorial advisory board member for the journal ONCOLOGY ®.

Transcript

Loaiza-Bonilla: Having this information come to us, particularly when it’s grounded in evidence, is key. For those who aren’t as familiar with RAG: it’s retrieval-augmented generation, which basically, after the model is trained, it’s able to draw on ground truths in terms of data — journals, as you mentioned, could be ASH [American Society of Hematology] or any of those sources — and integrate that into questions. The key is the prompt, coming from an expert like you. That’s why it’s so important on tumor boards to make sure it’s focused on the specific patient. I don’t think a lot of folks are doing this right now, in real time. Some may have software that does some of that, but not in a proactive manner. This is something folks listening and watching can probably take advantage of.

Matasar: I would encourage doing just that. Yes, you have to be cautious; not all systems are HIPAA compliant, so caveat emptor. We are oncologists, and we’re trained to use every tool at our disposal in the service of our patients. These are very powerful tools that are available in the here and now.

Loaiza-Bonilla: If you could look at all those tools, if someone is saying, “Okay, I want to try this out,” what do you feel has to be the value that those tools bring to them, so they can feel more confident about using them?

Matasar: The first thing to do is just take the leap. You don’t need to invest your decision-making into the LLM. You can just say, “Hey, this is what our tumor board generated out of a case — what would happen if I popped that clinical scenario into Doc GPT? What would it have advised us?” We’re interested in trying to run a prospective trial asking just this question: let’s run our tumor boards, and let’s also run more generative AI models that are free-thinking, like Claude or ChatGPT, whichever one you like, and let’s run RAGs as well, and see what those opinions are. You could then take those and offer them to external, blinded experts, and have them assess the case and the recommendations, and qualify those results. There are ways we can try to study this rigorously and prospectively. But for now, just start doing the work and get experience with what these models can and can’t do to support your practice — I think you’d be surprised.