Claude Opus 4.1 CA-EDA flags prior neurotoxic chemo as a pain-response predictor in celiac plexus radiosurgery

Conversational AI-assisted exploratory data analysis (CA-EDA) used three commercial LLMs (Claude Opus 4.1, ChatGPT 5.2, Gemini 3 Pro) to analyze raw individual patient data from a completed multicentre single-arm Phase 2 trial of celiac plexus radiosurgery (stereotactic radiation targeting the celiac plexus) for retroperitoneal pain syndrome in advanced cancer, largely pancreatic cancer (NCT03323489; n=125; 2165 variables). Claude completed a full statistical workflow in ~10–30 minutes, generated a visualization dashboard, and identified prior exposure to neurotoxic chemotherapy as a negative predictor of pain response (OR 0.20, p<0.001), yielding a 4-variable clinical score (AUC 0.71); ChatGPT produced a similar analysis but missed the neurotoxic-chemotherapy signal and built a 3-variable score, while Gemini produced narrative biologic hypotheses without statistical testing. Formal verification in Stata confirmed the key association (response 35% with prior neurotoxic chemotherapy vs 75% without; p<0.001), but because this was a post hoc predictor-finding exercise layered on a single-arm palliative RT trial, the practical “SOC benchmark” is less about ORR/PFS and more about whether the signal reproduces in external datasets and whether the score improves on clinician judgment alone.