GatorOnco matches oncologists on colorectal treatment accuracy

An AI-generated colorectal cancer treatment plan received almost the same correctness rating as one written by an oncologist. In a blinded, randomized evaluation by five UF Health oncologists, GatorOnco scored 4.09 for correctness while specialist plans scored 4.11, with no statistically significant difference (P=0.921). The paper posted August 10 is notable because it tests a high-stakes planning task—not merely medical question answering—requiring the system to combine patient details with clinical guidance that changes over time.

GatorOnco was developed with 282 billion tokens of biomedical text, including 166 billion tokens of health-system-scale clinical text from UF Health. Its domain-adaptation pipeline combines pre-training, model merging, two-stage post-training, and agent-based reinforcement learning. An agentic retrieval-augmented generation layer dynamically brings time-sensitive clinical guidelines into the reasoning process. That design addresses a practical problem in oncology: a fixed model can become stale as recommendations change, while treatment selection depends on staging, molecular markers, surgical eligibility, prior therapy, and other patient-specific evidence.

The evaluation separated several dimensions of plan quality. GatorOnco and the oncologists received statistically comparable scores for currency, at 4.04 versus 3.98 (P=0.478), and exactly the same 4.22 safety score (P=0.999). The model scored higher on readability, 4.46 versus 4.19, and completeness, 3.91 versus 3.52; both differences had P<0.01. The authors also report that GatorOnco significantly outperformed the open-source LLMs included in the comparison (P<0.01).

Those numbers do not establish that an AI oncologist has surpassed a human specialist. Five evaluators from one health system judged outputs produced in the same institutional setting. This was not a clinical trial that assigned treatment to patients and measured survival, adverse events, or quality of life. The scores reflect expert assessment of generated plans, and the paper does not establish that performance transfers to other hospitals, patient populations, or care systems. External validation matters especially because 166 billion training tokens came from the institution’s clinical text.

The study nevertheless moves the evaluation target in a useful direction. Instead of relying on a general medical exam score, it compares generated treatment plans with specialist work across correctness, currency, safety, readability, and completeness. Retrieving current guidance at inference time also offers a concrete way to narrow the gap between the pace of model retraining and the pace of clinical updates.

The next evidence should come from prospective, multi-center testing, subgroup-specific error analysis, traceable links between retrieved guidance and final recommendations, and disclosure of cases that required substantial specialist correction. For now, GatorOnco is not a license for autonomous care. It is evidence that an AI drafting system operating under specialist supervision may be ready for a more demanding stage of clinical validation.