The Oncologist Teaching AI to Understand Cancer
By Dr. Stephen Cole — 2025-10-20
When Dr. Stephen Cole first encountered AI in oncology, he recognized a pattern: systems that understood cancer statistics but not cancer patients. Sixteen years at a leading cancer center had taught him that the hardest decisions exist at the edge of human knowledge — and that's exactly where AI fails most dangerously.
The problem with oncology AI
"Oncology AI is proliferating fast," he says. "Treatment selection tools, prognosis models, imaging analysis systems. Some are genuinely useful. Some are dangerously overconfident. The difference usually comes down to whether real oncologists were involved in training them."
His IXO work focuses on treatment recommendation evaluation — testing whether AI outputs on complex cancer cases align with evidence-based guidelines while accounting for patient-specific factors. A model might correctly identify that a patient's tumor profile calls for immunotherapy while missing that their autoimmune history makes it contraindicated. Cole's annotations flag exactly these gaps.
The cases that matter most
"The hardest cases aren't the straightforward ones. They're the cases where the textbook answer is wrong for this particular patient. Teaching AI to recognize that requires someone who's lived through hundreds of those cases. You can't get that from data alone."
He also works on palliative care AI — a domain he considers critically undertrained. "Most medical AI training focuses on diagnosis and treatment. Almost none focuses on end-of-life communication, on how to talk to a family about stopping chemotherapy. That's a skill gap that matters enormously and almost nobody is addressing."
A message to fellow oncologists
"If AI is going to be used in cancer care — and it will be — the people who understand cancer have an obligation to be involved in building it. IXO makes that possible without asking you to become a data scientist."
Read The Oncologist Teaching AI to Understand Cancer on the IXO blog