US-China Medical Circles Shaken: Landmark Lung Cancer Trial Concludes as Medical AI Enters Real-World Practice

In a landmark study that pits artificial intelligence against hundreds of senior physicians, a Chinese medical AI model has demonstrated superior clinical decision-making in complex lung cancer cases—a result that signals a turning point for AI in real-world medicine.

Lung cancer remains the world's leading cause of cancer death. But the hardest cases for clinicians aren't the textbook examples—they're the "gray-area" patients where guidelines are unclear, expert opinions diverge, and every treatment decision carries risk.

That's precisely the challenge a team at Future Doctor, a Chinese health-tech company, set out to test. In what the researchers describe as the first large-scale empirical study in China to benchmark a medical AI against a cohort of over 100 physicians from top-tier hospitals, the results were striking: the company's medical large language model, MedGPT, outperformed 250 doctors from China's most prestigious hospitals in 88% of 40 real-world complex lung cancer cases.

The study, whose preprint was published on arXiv in September 2026, drew attention from nearly a dozen leading clinicians across China and the United States, including specialists from the Cancer Hospital of the Chinese Academy of Medical Sciences, West China Hospital of Sichuan University, UCSF Health, City of Hope National Medical Center, and the Miami CyberKnife Center.

From Passing Exams to Making Decisions

Large language models acing medical board exams no longer makes headlines. The gap between answering questions and practicing medicine, however, remains vast. In real clinical settings, the hardest decisions require weighing tumor stage, pathology, molecular markers, organ function, comorbidities, toxicity risks, patient preferences, and treatment sequencing—all at once, with no single correct answer.

The 40 cases in this study were drawn from actual clinical practice at top-tier Chinese hospitals and had never been publicly released. Most featured ambiguous guideline recommendations, conflicting comorbidities, or multiple intersecting factors that left the literature without clear answers.

"Real clinical cases don't follow textbooks," said Professor Liang, the study's principal investigator and a renowned thoracic surgeon. "Every patient has unique individual differences. The core challenge in the gray zone of clinical practice is how to apply evidence-based medicine while balancing safety with efficacy, and standardization with frontier approaches."

The 250 participating physicians all came from top-tier hospitals. Of these, 159—63.6%—were chief or associate chief physicians, the most senior ranks in China's medical hierarchy. They represented 27 provincial regions, 98 standardized centers, and three key specialties: thoracic surgery, respiratory medicine, and oncology.

The evaluation design was rigorous. Twenty senior chief physicians from six departments at elite institutions—including China-Japan Friendship Hospital and Peking University Cancer Hospital—developed 240 case-level evaluation criteria through three rounds of discussion before seeing any answers. Each case included a "safety red line": crossing it meant an automatic fail, testing whether a treatment plan properly accounted for risk, eligibility, intensity, monitoring, and fallback options.

Scoring was conducted blind to the source of each answer. Three AI judges scored independently based on expert-defined rules, with manual review by physicians. The study found that this scoring system aligned with doctors' judgments on the overall clinical acceptability of treatment plans.

Each doctor completed three cases under three conditions: unassisted, with search-based reference materials, and with MedGPT assistance. MedGPT independently completed all 40 cases under the same scoring rubric.

The result: in head-to-head case comparisons, MedGPT won 88 to 12 against the 250 physicians. Its standardized outputs were also more than 40% shorter than those of the doctors.

In a horizontal comparison with major international models, MedGPT scored 96.22 overall, compared with 71.38 for GPT-5.4, 61.81 for Claude Opus 4.7, and 54.62 for Gemini 3.1 Pro—suggesting that a specialized medical model trained for serious clinical decision-making holds a significant edge in the most difficult gray-area cases.

The More Important Number: 24.1%

But the 88-to-12 headline may not be the most consequential finding. The researchers point to another set of numbers as the real story.

Doctors working without assistance scored an average of 53.40. With search-based reference materials, that rose to 61.05. With MedGPT assistance, it climbed to 66.27—a 24.1% improvement over the unassisted baseline. For chief physicians specifically, the gain was even larger: 28.8%.

Notably, the search-reference condition and MedGPT drew from the same knowledge base. Simply giving doctors more information wasn't enough; organizing that information into a reviewable, structured treatment strategy made the difference. The improvement concentrated in three areas: candidate treatment pathways, information that must be gathered before a decision, and safety red lines with risk controls—precisely the elements most easily overlooked in complex cases.

MedGPT's approach mirrors this design. It lays out primary and alternative pathways, the conditions under which each path holds or should be abandoned, unknown information that could change the decision, and arrangements for re-evaluation and fallback. The resulting strategy document is handed to the physician or a multidisciplinary team for review, modification, and ultimately, the final call.

Professor Liang put it bluntly: "AI won't replace doctors. AI will replace doctors who don't use AI." He cautioned colleagues not to deify the technology—but also not to ignore it—envisioning an ideal future of tripartite decision-making among patients, physicians, and AI.

A New Era Arrives

"AI beats doctors" proves a technological breakthrough. "AI augments doctors" determines the value of medical AI. That distinction is central to Future Doctor's thesis on medical artificial general intelligence: AI won't replace physicians, but it will redefine the boundaries of their capabilities, giving more doctors access to decision-making quality that was once the domain of a small elite.

Wang Shirui, founder and CEO of Future Doctor, frames the mission in terms borrowed from the 11th-century Chinese philosopher Zhang Zai. "Preserving lost knowledge" means standardizing and scaling the diagnostic capabilities of today's top doctors, breaking the impossible triangle of quality, affordability, and accessibility in healthcare. "Opening peace" means AI doesn't just replicate existing answers—it helps create new medical knowledge, compressing the typical three-to-five-year clinical guideline update cycle into a quarterly rhythm.

The company's trajectory has been methodical: 96% diagnostic consistency with attending physicians at West China Hospital in 2023; the CSEDB clinical safety-efficacy dual-track benchmark, developed with 32 experts and published in npj Digital Medicine in 2025; and now, this head-to-head trial against hundreds of top-tier physicians on the hardest gray-area decisions.

International reaction has been notably positive. Dr. Robert Warren, director of surgical oncology at UCSF Health, called it "an important study" on X, noting that existing data point toward AI playing a prominent role in diagnosis and treatment decisions. Dr. Hope S. Rugo, director of breast medical oncology at City of Hope—an NCI-designated comprehensive cancer center of the highest rank—described it as "a truly remarkable advance in oncology" and expressed interest in collaboration. Dr. Mark Pomper, medical director of the Miami CyberKnife Center, said AI is transforming cancer treatment and the next step is collaborating with more oncologists to scale patient access. Dr. Marc Shuman, professor emeritus of hematology-oncology at UCSF, expressed hope that the research would bring AI into real-world practice.

From thoracic surgery to breast oncology to radiation oncology, clinicians in both countries focused on the same question: not how high AI can score, but whether it can enter real clinical workflows and help doctors make complex decisions more completely and more safely.

AlphaFold cracked protein structures. AI drug discovery has compressed development timelines. But these breakthroughs all must pass through the final piece of the puzzle—clinical decision-making—before they truly reach patients. This study offers the first large-scale, human-controlled empirical evidence that medical AI is crossing the watershed from answering questions to making complex decisions.

The research team has confirmed that the next phase will embed these capabilities into real clinical workflows, with regulatory registration completed before deployment, to measure actual impact on treatment processes and patient outcomes.

Future Doctor will also launch an academic empowerment initiative for oncologists called the "Oncology AI Diagnosis and Treatment Capability Enhancement Project," to be unveiled at the 2026 China Integrated Oncology Conference. The program will use timed online case responses focused on gray-area cases without standard guideline answers, with human-AI collaboration to determine optimal treatment pathways. AI will score decisions across six quality dimensions. The project includes mid-term and summary meetings with on-site defenses, and participating physicians at the summary stage will be eligible to co-create research projects worth tens of millions in funding.

The goal, the company says, is to make MedGPT a clinical partner for physicians—not a replacement, but a collaborator working side by side.