(Yicai) Sept 28 -- Digital twins are finding their way into clinical trials as drugmakers explore their potential to predict treatment outcomes, though wider adoption will depend on sufficient patient data and regulatory acceptance.

In the past month, Chen Huan, founder of a medical modeling company, told Health Insight that the firm has received requests from four groups of clients, including pharmaceutical companies, contract research organizations, and hospitals, seeking artificial intelligence-powered virtual patients that can predict drug effects in early-stage clinical trials.

AI virtual patients have traditionally been used in medical education, but advances in large language models are expanding their use into clinical research. The technology can create a digital twin from a real patient's health data to simulate how their condition may evolve under different treatments, potentially improving clinical decisions and reducing research costs.

But wider use still faces hurdles. Limited patient data can make reliable models difficult to train, while generative AI's black-box nature makes its results difficult to interpret and trust.

Still, regulatory support is growing. In June, the US Food and Drug Administration issued guidelines allowing evidence from model-informed drug development to provide supplementary support for regulatory decisions. Five days later, China's National Medical Products Administration and three other government departments encouraged the use of virtual patients, or digital twins, to simulate and optimize treatment plans before real-world trials.

"This indicates a shift in regulatory consensus; digital twins are transitioning from being a single-point modeling tool to an early clinical decision support system," Chen said, adding that the shift is behind the recent surge in client demand. Virtual patients have had clear potential buyers from the outset, he added, as the technology ultimately targets pharmaceutical companies' clinical development budgets.

The technology is already drawing interest from hospitals. Qilu Hospital of Shandong University recently announced plans to procure a "digital twin virtual clinical trial" system for virtual population modeling and efficacy prediction. The project has a budget of CNY2.6 million (USD387,300).

China alone had nearly 3,000 new drug clinical trials in 2025. If just 20 percent of trial projects eventually use the technology and each project generates CNY2 million in service fees, the potential market would exceed CNY1 billion (USD149 million).

From Control Groups to Digital Twins

The approach differs fundamentally from conventional clinical trials.

Traditional clinical trials compare similar patient groups receiving different treatments, while digital twins seek to predict how a virtual version of the same patient would respond, according to Xu Liang, a senior AI pharmaceutical industry professional.

Patient information is often fragmented and stored in different formats, including medical records, pathology reports, imaging, genetic data, time series, and doctors' notes, making it difficult to analyze collectively. Xu said that LLMs are beginning to map this diverse information into a unified physiological trajectory for individual patients.

"At this point, the technology side has at least been able to make this work," the medical expert noted.

Data Hurdles Remain

Technical feasibility, however, does not mean the technology is suitable for every disease.

Digital twin technology must first identify suitable clinical trial scenarios if it is to advance further, according to Chen from the medical modeling company. Early research has focused on neurodegenerative diseases such as Alzheimer's, Parkinson's, and Huntington's disease because they have clear longitudinal scales, continuous disease progression, and extensive historical clinical trial data.

By contrast, conditions are not yet mature for applying digital twins to rare diseases and cancers, despite strong demand for the technology, Chen said. Rare diseases in particular could benefit from reducing reliance on control groups but face a data paradox.

"Because there are few patients, we need digital twins; however, conversely, the scarcity of patients means there is limited data available, making it difficult to train reliable models," Chen said.

Generative AI also operates as a black box, making its results difficult to interpret and trust.

"Data determines whether a model can be developed, while regulation determines how much that model will ultimately be valued," Xu said, pointing to the two factors that will shape their wider use.

Editor: Emmi Laine