Collaborative AI system helps accelerate clinical trial design

An artificial intelligence system that operates like a collaborative team of medical experts could accelerate clinical trial design, one of the most difficult steps in drug development, according to a new study by Weill Cornell Medicine investigators. The findings, published July 7 in Nature Communications, evaluated the potential of the system, called EmulatRx, to simulate, design and improve clinical trials using real-world patient data. A necessary step before a drug is approved for market, a randomized clinical trial is a rigorous study in which participants are randomly assigned into separate groups to evaluate how well treatments or interventions work and their side effects.

The researchers predict that combining real-world data with collaborative AI reasoning has the potential to make clinical trials faster, more affordable and more precise, which could ultimately lead to new treatments for patients with higher success rates.

Randomized clinical trials are the gold standard for determining whether new treatments are safe and effective. But designing them is a slow, expensive and highly complex process."

Dr. Fei Wang, senior author, associate dean for AI and data science and the Frances and John Loeb Professor of Medical Informatics, Department of Population Health Sciences, Weill Cornell Medicine

 Researchers must decide who qualifies to participate, how treatments are compared, what outcomes to track and whether enough patients can be recruited. These decisions typically require intensive collaboration among trialists, clinicians, statisticians and data specialists.

Dr. Wang's team built EmulatRx to streamline this decision-making process into one system that can exchange information, identify problems and revise recommendations for a clinical trial.

Drs. Haoyang Li, Weishen Pan and Chengxi Zang of the Institute of AI for Digital Health in the Department of Population Health Sciences at Weill Cornell, and Dr. Suraj Rajendran, a graduate of the Tri-Institutional Computational Biology and Medicine Program, all co-authors of the paper, contributed to the research.

A virtual research team

EmulatRx is organized around five specialized computational agents that mirror a scientific team. "Because each EmulatRx agent is empowered by a large language model, they can exchange information in natural language and work together much as human experts do," Dr. Wang said.

At the center of the system is a coordinating "Supervisor" agent that manages workflow and integrates outputs. A "Trialist" reviews past studies and extracts key elements—like eligibility criteria, treatments and outcomes—to outline a trial structure. An "Informatician" translates those requirements into queries that can identify appropriate patients in real-world data such as electronic health records. A "Clinician" ensures the design makes medical sense and references published research, while a "Statistician" evaluates potential outcomes and estimates how a treatment might perform using real-world data.

Learning from real-world patients

To evaluate EmulatRx, the researchers used de-identified electronic health records from large clinical databases covering both acute conditions (heart failure, septic shock, kidney injury) and chronic diseases (Alzheimer's and Parkinson's). These records included diverse populations—such as older adults or patients with multiple conditions—who are often underrepresented in traditional trials.

EmulatRx analyzed the data through "target trial emulation," which applies key features of a randomized clinical trial—eligibility criteria, treatment groups, follow-up and outcomes, and causal contrast—to information collected during routine care. This included searching free-text clinical notes alongside structured information such as diagnosis codes, medications and laboratory results.

Using this information, the system identified appropriate patient groups for replicating known findings and investigating differences in treatment effects across patient subgroups that may not have been apparent in the original trials. For example, the system flagged when a treatment benefited one group but posed risks to another, helping researchers design more precise and safer trials from the outset.

Across historical clinical trials, the system reproduced many previously reported treatment effects, suggesting it could help researchers evaluate trial designs before launching expensive studies.

Human in the loop

A central feature of EmulatRx is the ability for researchers to monitor and intervene in its work. They can follow the agents' exchanges and review each stage of the analysis, while an expert can pause the process, correct a decision or direct the system to reconsider its approach. EmulatRx also learns from those corrections, reducing the chance of repeating mistakes.

"It is important to keep a human in the loop, so the system does not go in an unreasonable direction," Dr. Wang said.

Before EmulatRx is ready for clinical or commercial deployment, it will require broader validation across other health systems and types of patient data. The researchers are working on commercial development and hope to make it available for investigator-initiated trials at universities as well.

"We still need randomized controlled trials," Dr. Wang said. "The question is how to design them, so they can be conducted more efficiently and have a higher chance of success."