Future Doctor's MedGPT beats 250 hospital physicians 88-12 in complex lung cancer test

Chinese medical AI company Future Doctor has published results from a large clinical evaluation of MedGPT, its medical large model, on complex lung cancer cases. Across 40 unpublished cases, the model's treatment strategies beat those of 250 physicians from top-tier Chinese hospitals by 88 to 12 in blinded comparisons.

The study was run with thoracic surgery teams at leading tertiary hospitals and deliberately targeted the clinical gray zone, where guideline recommendations are unclear, comorbidities conflict, and no single answer is standard. The 250 participating physicians came from 27 provincial regions and 98 standardized centers, spanning thoracic surgery, respiratory medicine, and oncology; 159 of them, or 63.6%, were chief or associate chief physicians.

Each doctor handled three different cases under three conditions — unassisted, with retrieval references, and with MedGPT support — producing 2,250 strategy documents that allowed the team to compare solo judgment with human-AI collaboration.

The results have two parts. Among 750 unassisted physician plans, MedGPT received the higher score 660 times, and its standardized output was on average more than 40% shorter. When doctors worked with MedGPT, their overall scores rose from 53.40 to 66.27, a 24.1% relative gain.

In a same-condition comparison with general-purpose models, MedGPT scored 96.22, ahead of GPT-5.4 at 71.38, Claude Opus 4.7 at 61.81, and Gemini 3.1 Pro at 54.62. The article attributes much of the gap to candidate path generation: general models tend to propose a single plan, while MedGPT lays out multiple options with entry and exit conditions for each.

To make the grading defensible, 20 multidisciplinary experts spent three rounds building 240 case evaluation rules, 40 of them hard safety boundaries around treatment eligibility and intensity that trigger an automatic failure if crossed. After the rules were frozen, three AI judges scored the outputs independently with the source of each answer hidden.

MedGPT 2.0 runs a multi-specialty agent consultation, compiles the case timeline, generates a standard E-Path and a frontier F-Path, then retrieves missing evidence and simulates outcomes before converging on a V-Path verification plan. Future Doctor says its lung cancer evidence base holds about 13,000 source documents organized into 201,000 clinical evidence units.

Clinicians including Li Yin of the Chinese Academy of Medical Sciences Cancer Hospital, Wang Yongsheng of Shandong Cancer Hospital, Chen Longqi of West China Hospital, and UCSF Health's Robert Warren and City of Hope's Hope S. Rugo endorsed the findings. The company says the next step is validating the approach across more clinical tasks.

The caveat is that this is a company-led study on one cancer type and a specific case set, however large the panel. It shows a medical model can produce structured, checkable options in gray-zone decisions; it does not yet show independent reproducibility or a change in clinical guidelines, which is what to watch next.