Virtual Clinical Trial Emulation with Generative AI Models

This research will apply newly emerging generative AI technology to transform biomedical and health research by enabling virtual clinical trial emulation with synthetic data. It will overcome key limitations in both Randomised Controlled Trials (RCTs) and observational studies. RCTs have long been considered as the "gold standard" to evaluate treatments and medicines. However, they are far from being able to answer all clinical questions. In addition to time and cost constraints, RCTs have significant limitations to generalise their findings as their scope is limited. In many situations conducting RCTs with real patients is logistically challenging or unethical due to their potentially harmful nature. This leaves a significant knowledge gap, for example, we still have very limited clinical guidelines about how to manage multi-morbidities. While observational studies can overcome the issues faced by RCTs by leveraging routinely collected data from real world, they are typically imbalanced across population, diseases and interventions; there are a significant amount of noise and missing measurements in the data, and we need lengthy time and significant effort to remove patient identifiable information from the data to protect privacy. More importantly, treatment choices and outcomes in real world clinical cases may depend on factors that are not measured within the data, which may invalidate the observational study.
AI research has made great advances in creating new data. With new generative AI models, we can generate synthetic patient populations that faithfully preserve the statistical attributes of real populations. Compared with anonymised real data, synthetic data can be generated in unlimited volume while containing "zero" information about real individuals. Hence, they are in a much better position to overcome legal barriers in data protection and sharing. More importantly, experiments with synthetic data will allow clinical researchers to perform "virtual-trials" to gain quantitative insight into causal relations between treatment and its effect. This will enable prediction and comparison of hypothetical treatments to seek answers to important research questions that currently cannot be answered in real trials. The success of this adventurous and timely research will bring a landscape change to revolutionise future biomedical and health research by broadening its research agenda, liberating its restrictions, saving cost and time, leading to significant benefits to healthcare by speeding up new timelines for treatment discovery, addressing increasingly complex healthcare landscape in elderly population and multi-morbidity, and transforming regulatory and policy making process.

This research will apply newly emerging generative AI technology to transform biomedical and health research by enabling virtual clinical trial emulations with synthetic data. AI research has made great advances in creating new data. With new generative AI models, we can generate synthetic patient populations that faithfully preserve the statistical attributes of real populations. Experiments with synthetic data will allow clinical researchers to perform "virtual-trials" to gain quantitative insight into causal relations between a target effect and its causes. This will enable prediction and comparison of hypothetical interventions to seek answers to important research questions that currently cannot be answered in real trials, including clinical questions associated with underrepresented populations of children, older adults, and patients with multi-morbidities and polypharmacy - these people are commonly excluded in trials.

The project aims to validate the concept by seeking answers to the primary research questions including: 1) Can we use generative AI models to generate synthetic data that preserve the same value for research as real-world health data? 2)Can we perform virtual clinical trial emulations by discovering correct causal relations from the synthetic data?

Within the scope of this feasibility study, we will focus on a specific Type 2 diabetes mellites (T2DM) use case, which is a confirmatory study to test the ability of the AI-driven virtual trial emulations by emulating established clinical trials to replicate their results. We will target the effect of an exemplar drug (i.e. Liraglutide, a GLP-1 receptor agonist). Over the years, several RCTs have been conducted in the LEAD (Liraglutide Effect and Action in Diabetes) program to collect detailed evidence about the effect of this drug on HbA1c and weight, cardiovascular risk, the lipid profile, and blood pressure. The trial emulations will attempt to rediscover their findings.