Decision-only medical AI: early evidence and pairing with language models

Omar, Ramaswamy, Sakhuja and other researchers tested two decision-focused medical AIs, Jev and OpenAI Decisions, and the large language model GPT-6 Luna across 2,134 case presentations and 2,483,222 model calls. The paper says that when labels or pressure prompts were added without changing the clinical findings, the decision models changed 13.4 more decisions per 100 decisions than under neutral wording, versus 14.0 for the language model; after pairing the decision model with the language model, this figure fell to 2.6, a reduction of about 81% on the development set and 42% to 48% on the held-out reports. Triage accuracy remained around 81% to 82%. The authors say the pairing pipeline costs at most $0.77 per 1,000 records and describe it as a practical proof of concept for reducing framing sensitivity and flagging text that needs review.